A Computational Study of Lesion Morphology Influence on Fractional Flow Reserve in Coronary Artery Disease
Students: Turash Asif Ahmed
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Clinical evaluation of coronary artery disease is shifting from anatomical observation toward functional characterization. While Fractional Flow Reserve (FFR) is a definitive indicator of myocardial ischemia, conventional transient CFD simulations are computationally expensive for rapid clinical use. Building upon a resistance-based steady-state framework, this study investigates plaque morphology’s influence on hemodynamics within the Left Main Coronary Artery (LMCA). Patient-specific geometries were reconstructed from CTA data using Materialise Mimics and 3matic before simulation in ANSYS Fluent. Microvascular outlet resistances were incorporated to simulate hyperemic conditions. This framework enabled systematic investigation into how lesion morphology influences pressure distribution beyond conventional diameter reduction. Results demonstrate that while stenosis severity is the primary driver of pressure loss, lesion length and plaque eccentricity act as significant physiological multipliers. Asymmetric and elongated plaques produced more profound pressure drops than symmetric stenoses of equal severity. While this study is limited by the assumption of rigid arterial walls and steady-state flow, it provides a rapid, non-invasive tool for functional assessment, highlighting the necessity of incorporating complex lesion geometry into patient-specific diagnostic planning.
A Multimodal Framework for Mpox Screening: Lesion-Focused Image Analysis and Clinical Feature Fusion
Students: Nafisa Ferdous
Supervisor: Dr. Taufiq Hasan, Professor
Abstract:
Automated screening of Mpox remains challenging due to the limitations of image-only deep learning models, which often fail to capture clinical context and are sensitive to background noise, multiple lesions, and variability in image quality. This study proposes a multimodal machine learning framework that integrates skin lesion images with clinical metadata to improve patient-level Mpox classification, particularly in resource-constrained settings. The study utilizes the Uvira Mpox dataset, comprising metadata records of 939 patients, of which 154 patients had both PCR-confirmed labels and usable images for model development. A pretrained EfficientNet-B0 model trained on full images achieved moderate performance (AUC ≈ 0.69), highlighting the limitations of whole-image analysis. A metadata-based CatBoost model with SHAP-based feature selection achieved stronger performance (AUC ≈ 0.75). Late fusion of image and metadata predictions further improved results (AUC ≈ 0.74-0.75). To enhance image-based learning, a lesion-centric preprocessing pipeline, including annotation, bounding box merging, adaptive cropping, and Non-Maximum Suppression was introduced, generating over 7,000 lesion crops. This significantly improved image-only performance (AUC ≈ 0.84) and boosted fusion results. Additional analysis across CT value groups revealed variability with viral load, while low correlation between modalities supported the effectiveness of fusion. Overall, this work demonstrates that lesion-focused preprocessing and multimodal fusion significantly improve Mpox screening, despite limitations such as small dataset size.
A simple and Rapid PEG-Coated Silver Nanoparticle-Based Colorimetric Sensor for Salivary Creatinine Detection
Students: Afroza Yesmin Isha
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Creatinine is a vital biomarker for assessing kidney function and early diagnosis of chronic kidney disease (CKD), and its detection in saliva offers a promising non-invasive approach for point-ofcare (POC) applications. In this study, a simple, rapid, and cost-effective colorimetric sensing platform was developed based on polyethylene glycol (PEG)-coated silver nanoparticles (AgNPs) for salivary creatinine detection. The sensing mechanism relies on creatinine-induced aggregation of PEG-coated AgNPs, resulting in a visible color change accompanied by a shift in surface plasmon resonance (SPR) absorbance. Experimental conditions were optimized to improve sensitivity, stability, and reproducibility. The developed sensor demonstrated a reliable analytical response within the relevant concentration range, indicating its suitability for detecting low concentrations of creatinine in saliva. The PEG-coated AgNPs exhibit good colloidal stability, reduced non-specific interactions, and enhanced biocompatibility while maintaining low material and preparation costs. The proposed method offers key advantages, including rapid response, minimal sample preparation, and the potential for visual detection without sophisticated instrumentation. These features make it highly suitable for resource-limited settings and support its potential for future development as a portable POC diagnostic tool for CKD screening and monitoring.
A Weakly Supervised Reciprocal Attention-Guided Network for Bias Mitigation in Late-Gestation Dating and Adverse Pregnancy Outcome Prediction
Students: Md. Mehedi Hassan, Hasib Al Siam
Supervisor: Dr. Taufiq Hasan, Professor
Abstract:
Precise late-gestation dating and the prediction of adverse pregnancy outcomes are critical for maternal-fetal triage. While deep learning-based approaches have the potential to address these challenges, the “Clever Hans” effect remains a prevalent risk, wherein models learn to exploit ultrasound interface artifacts rather than true fetal biology. To accurately predict the lateterm gestational age and adverse pregnancy outcomes accurately, we developed a multimodal clinical decision support tool (CDST) utilizing data from the Projahnmo Research Foundation (PRF) branch study and the WHO AMANHI cohort. Gestational age labels were anchored to goldstandard early ultrasound ground truth. We mitigated artifact bias using defensive preprocessing and weakly supervised object localization (WSOL), where a pristine Zenodo dataset guided fetal localization in unannotated clinical images. Visual features were extracted using an EfficientNetB4 backbone and fused with clinical biometrics via a novel Reciprocal Attention-Guided Discrepancy Gate (RADGate). On a strictly locked blind test cohort, the proposed model achieved a mean absolute error of 5.31 days and a strong correlation of 0.9714 in the late-pregnancy gestational age prediction. For the prediction of adverse outcomes, the proposed CDST yielded a strong AUC of 0.81. Furthermore, attention saliency maps confirm effective mitigation of artifact bias, providing crucial explainability. By integrating WSOL with a defensive processing pipeline, the proposed model effectively mitigates artifact-driven shortcut learning. Consequently, this multimodal CDST provides a highly robust framework for fetal assessment, bridging the gap between advanced deep learning and the practical needs of resource-constrained clinical environments. Keywords – Reciprocal Cross-Attention; Fetal Head Biometry; Gestational Age; WSOL; Clever Hans Effect; CDST.
Anatomical and Biomechanical Assessment of Left Coronary Artery Plaque Vulnerability Using CCTA-Based One-Way FSI
Students: Mehedi Hasan Nirzan
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Background: Coronary plaque rupture is a leading cause of major adverse cardiovascular events (MACEs), and its assessment requires understanding both anatomical and biomechanical factors. Objective: This study aims to evaluate biomechanical and anatomical indicators associated with plaque vulnerability in the left coronary artery using patient-specific coronary CT angiography (CCTA) and one-way fluid-structure interaction (FSI) simulations. Methods: Patient-derived coronary geometries, including lumen, arterial wall, and plaque regions, were reconstructed from CCTA data using Materialise Mimics and Materialise 3-matic, followed by discretization for computational analysis. Hemodynamic simulations were performed in ANSYS Fluent to obtain pressure and wall shear stress distributions, which were then transferred to Transient Structural for analysis of arterial wall and plaque behavior using a one-way FSI approach. Key biomechanical parameters, such as peak cap stress and flow-derived indices, were computed as potential markers of plaque instability. Results: Preliminary results are expected to show non-uniform wall shear stress and pressure, with higher stresses near plaque regions. One-way FSI may indicate elevated peak cap stress in stenotic plaques. Low WSS and high gradients are likely associated with vulnerable areas. Conclusion: The proposed framework offers a feasible approach for integrating medical imaging and computational modeling to assess biomechanical factors linked to plaque vulnerability. This methodology may contribute to improved risk stratification by identifying mechanical indicators associated with increased likelihood of plaque rupture.
Antimicrobial Silver Nanoparticle-Embedded Decellularized Amniotic Membrane Powder for Burn Wound Healing
Students: Ishtiaq Reza
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Burn injuries remain a significant global health challenge, particularly in developing countries like Bangladesh, where limited resources and delayed access to advanced care often worsen patient outcomes. Conventional burn wound dressings are often inadequate for managing deep or chronic wounds, leading to delayed healing, increased risk of infection, and permanent scarring. The human amniotic membrane (HAM), a naturally derived biomaterial, exhibits remarkable antiinflammatory, anti-fibrotic, and regenerative properties, making it highly suitable for wound healing applications. Decellularization of HAM preserves its extracellular matrix (ECM) architecture while eliminating immunogenic components, enabling the development of highly biocompatible scaffolds. In this study, decellularized HAM (d-HAM) was processed into a fine powder form to enhance stability, storage, and ease of clinical application. To effectively combat infection-a major complication in burn wounds-silver nanoparticles (AgNPs) were incorporated into the d-HAM powder due to their potent and broad-spectrum antimicrobial activity. AgNPs exert their effects through multiple mechanisms, including disruption of microbial cell membranes, generation of reactive oxygen species (ROS), and interference with DNA replication, thereby minimizing the risk of microbial resistance. Their nanoscale size allows enhanced surface interaction and sustained release of silver ions, ensuring prolonged antimicrobial efficacy. The composite material was characterized through physicochemical and biocompatibility analyses. Findings indicate that AgNP-integrated d-HAM powder significantly reduces microbial load, accelerates epithelialization, and promotes collagen deposition, demonstrating its strong potential as a safe, effective, and advanced antimicrobial therapy for burn wound management.
BELKA BERT: A Robust Multimodal Molecular Representation-Based Framework for Multitarget Drug Target Prediction
Students: Mehnush Morshed, Nufayer Jahan Reza
Supervisor: Dr. Jahid Ferdous, Associate Professor
Abstract:
Early-stage drug discovery increasingly relies on computational models to identify molecules that bind multiple protein targets because large-scale experimental screening is costly and time-intensive. Although recent representation learning methods have improved molecular prediction, many studies still depend on a single chemical view and are evaluated under settings that do not fully reflect realistic generalization. Motivated by this gap, this study proposes BELKA BERT, a leakage-aware information fusion framework for multilabel binding prediction that integrates three complementary molecular representations: a ChemBERTa encoder over SMILES, Morgan fingerprint descriptors, and learnable embeddings of three synthesis building blocks. Developed on 1.61 million compounds from the Leash BELKA dataset for prediction against BRD4, HSA, and sEH, the framework uses label combination stratification with building block triplet grouping to remove synthesis group overlap between training and test data and enable more realistic evaluation. The learning pipeline further combines domain adaptive masked language modeling, supervised fusion learning, class imbalance correction, promiscuity regularization, temperature scaling, and target-specific threshold optimization to improve ranking performance and decision reliability. On the leakage safe test setting, the final fusion model achieves a micro average precision of 0.972 and a macro average precision of 0.952. In aligned ablation experiments, the full fusion architecture improves micro average precision from 0.934 to 0.953 and macro average precision from 0.906 to 0.929 versus a sequence-only variant, while the largest drop occurs when fingerprint features are removed, confirming strong complementary value beyond transformer sequence embeddings and supporting trustworthy multitarget virtual screening. 1 Figure 1: Overall workflow of BELKA-BERT 2
Calcium-Activated Chitosan/Alginate Layer-by-Layer Engineered Hemostatic Cotton Gauze for Accelerated Coagulation and Clot Stabilization
Students: Nafisa Akter
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Hemorrhage due to trauma causes 30% to 40% of trauma related death worldwide and is one of the leading causes of death for both military and civilian traumas. Existing hemostatic agents focus primarily on achieving rapid clot formation ignoring clot stabilization and prevention of fibrinolysis. In this study, a multi-functional hemostatic gauze was developed using woven cotton which was functionalized by layer-by-layer (LbL) assembly of chitosan and sodium alginate. Calcium chloride was incorporated into the system to facilitate intrinsic hemostatic activity and tranexamic acid (TXA), a potent antifibrinolytic agent, was added to inhibit fibrinolysis and stabilize the formed blood clot. The engineered gauze demonstrated excellent biocompatibility and strong pro-coagulant capabilities as evidenced by a very low hemolysis value (~0.06% ± 0.03%) and an average blood clotting index (~14.5 ± 0.5%) compared to commercially available gauze. In the mice liver laceration model, the gauze achieved hemostasis approximately 10 seconds earlier than kaolin-based commercial hemostatic gauze. Similarly, in the rat tail amputation injury model, the gauze achieved hemostasis much earlier than the untreated gauze demonstrating improved hemostatic efficiency. The combination of a mechanically-robust cotton support with an effective procoagulant-antifibrinolytic treatment that works together to provide a scalable, effective method to quickly stop bleeding and will provide additional options for trauma management and administering care in emergency situations or surgery in under-resourced communities.
Comparative Analysis and Optimization of Detergent and Salt-Based Decellularization Protocols for Functional Bovine Pericardial Scaffolds
Students: Mahzabin Afroz Shithi
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Decellularized extracellular matrix (dECM) scaffolds derived from xenogeneic tissues have emerged as promising biomaterials in tissue engineering due to their structural similarity to native tissue and inherent bioactivity. This study focuses on the optimization of chemical decellularization protocols for bovine pericardium using detergent- and salt-based approaches, with particular emphasis on effective DNA removal as a key indicator of decellularization efficiency. Initially, conventional protocols employing sodium dodecyl sulfate (SDS, 1%) and Triton X-100 (1%) at varying exposure durations (5-24 hours), as well as combined SDS-Triton treatments, were evaluated. To improve extracellular matrix (ECM) preservation while ensuring sufficient cellular removal, modified protocols were developed, including low-concentration SDS (0.1%), Triton X-100 (0.5%), sequential treatment with 2 M NaCl followed by Triton, and prolonged exposure to 0.24% SDS. The effectiveness of decellularization was comprehensively assessed through histological analysis, tensile testing, hemocompatibility assays, and Attenuated Total Reflectance-Fourier Transform Infrared (ATR-FTIR) spectroscopy. Critically, quantitative DNA analysis was performed to determine residual nucleic acid content, which serves as the primary criterion for evaluating decellularization success. The goal was to achieve DNA levels below acceptable thresholds while maintaining ECM integrity. Preliminary findings indicate that reduced detergent concentrations combined with salt-based treatments significantly enhance DNA removal while minimizing structural and biochemical damage to the scaffold. This study establishes correlations between detergent concentration, exposure duration, and DNA content, providing insight into optimizing decellularization protocols. Overall, the optimized strategy aims to produce biocompatible, mechanically stable, and low-immunogenic scaffolds suitable for regenerative medicine applications, particularly in soft tissue engineering.
Design and Physicochemical Characterization of Polysaccharide-Functionalized Magnetic Nanoparticles for Targeted Anticancer Drug Delivery
Students: Nafish Ahanaf
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Bare iron oxide nanoparticles (IONPs) face significant limitations in biomedical applications due to colloidal instability, aggregation, and rapid systemic clearance. This study investigates the surface engineering of IONPs using three distinct biodegradable polysaccharides – dextran, chitosan, and alginate to develop highly stable, efficient nanocarriers for the targeted delivery of the anticancer drug doxorubicin (DOX). Paramagnetic IONPs were functionalized via physical adsorption for dextran, and electrostatic ionic gelation for chitosan and alginate, using sodium tripolyphosphate (TPP) and calcium chloride as crosslinkers, respectively. The successful surface modification of the magnetic cores was confirmed via FourierTransform Infrared (FTIR) spectroscopy, which revealed the preservation of the characteristic Fe-O stretch alongside the emergence of distinct polymeric functional groups, including glycosidic linkages (C-O-C), amine bending (N-H), and carboxylate stretching (-COO⁻). To evaluate the therapeutic potential of these functionalized nanocarriers, DOX was encapsulated within the polymeric shells. The in vitro drug release kinetics were subsequently assessed utilizing a dialysis membrane model to determine the bulk cumulative release profiles and evaluate the controlled-release capabilities of each coating. Comprehensive physicochemical characterization, including Dynamic Light Scattering (DLS), Transmission Electron Microscopy (TEM) is employed to correlate the hydrodynamic diameter, surface charge, and surface morphology with the observed drug release kinetics. Ultimately, this comparative investigation aims to identify the optimal polysaccharide architecture for the sustained delivery of DOX, advancing the development of magnetic, biocompatible theranostic agents for targeted cancer therapy.
Development and Evaluation of an Artificial Intelligence Model for Cervical Cancer Detection from Colposcopic Images
Students: Nuzhat Aisha Shaikh, Dania Khan
Supervisor: Dr. Taufiq Hasan, Professor
Abstract:
To address the global health crisis of cervical cancer, particularly in low- and middle-income countries like Bangladesh where trained colposcopists are scarce, we developed a deep learning framework designed for accessible nationwide screening through automated CIN grading and Swede score prediction. Grounded in a novel dataset from Bangladeshi Hospital combined with IARC data, our dual-pathway architecture utilizes a hybrid ensemble for three-class diagnostic grading (Normal vs. CIN1 vs. High-Grade) and a dual-branch system that replicates clinical logic by predicting individual Swede score components.The results demonstrate a 72% diagnostic accuracy for CIN grading. Additionally, the approach showed high reliability in evaluating Swede score components, achieving 72% accuracy for acetic acid uptake and 79% accuracy for iodine staining.This methodology performs colposcopic assessment, reduces diagnostic subjectivity, and provides a scalable solution to support the WHO’s cervical cancer elimination targets in resource-limited settings.
Development and Evaluation of Eucalyptus Oil-Based Diclofenac Sodium Nanoemulgel for Controlled Topical Drug Delivery
Students: Mst. Tasnim Fariha Khanom
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Pain and inflammation are prevalent clinical challenges commonly managed using topical nonsteroidal anti-inflammatory drugs (NSAIDs) due to their reduced systemic effects. Among NSAIDs, diclofenac sodium is widely employed due to its favorable efficacy. This study aimed to develop a nanoemulsion-based gel (nanoemulgel) for controlled topical delivery of diclofenac sodium, using eucalyptus oil as the oil phase and Tween-80 as the surfactant, subsequently incorporated into a gel base. The optimized nanoemulsion exhibited an acceptable droplet size distribution with a polydispersity index (PDI) of 0.49 ± 0.05. In vitro drug release studies demonstrated controlled and sustained release of 95% over 12 days. The nanoemulgel demonstrated excellent spreadability, a favorable zeta potential indicative of colloidal stability, and a skin-compatible pH of approximately 6. No allergic reactions or irritation were observed on rabbit skin following 72-hour application. Ex vivo skin permeation studies conducted on excised mouse skin demonstrated that the nanoemulgel achieved a 5-fold higher permeation flux compared to a commercially available diclofenac gel over 48 hours, suggesting superior topical delivery. These findings indicate that the developed nanoemulgel represents a promising controlled drug delivery platform for the management of pain conditions, including arthritis, with potential for improved patient compliance and prolonged therapeutic effect.
Development and Evaluation of Gum Based PEGylated Chitosan Nanogels for Platelet Lysate Delivery in Burn Wound Regeneration
Students: Jannatul Ferdusy Saily
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Burn Wound management remains a significant clinical challenge due to its high susceptibility to infection, impaired vascularization and reduced availability of endogenous growth factors. Standard treatment approaches predominantly target microbial control. However, they exhibit limited regenerative efficacy and occasionally induce cytotoxicity and antibiotic resistance. The present study aimed to engineer a non-invasive, biocompatible nanogel system to enhance burn wound regeneration using platelet lysate (PL) incorporated within gum based natural polymeric nanoparticle (NP) systems. Chitosan NPs were synthesized via ionic gelation utilizing natural gum as crosslinking agents, followed by surface PEGylation using musselinspired catechol chemistry to augment physiochemical stability. The resulting NPs were subsequently embedded within a carbomer matrix to yield topical nanogel composition. Two distinct formulations- gum Arabic and xanthan gum were systemically evaluated through in vitro release kinetics and in vivo assessment in a second degree burn wound mice model. The gum Arabic based formulation exhibited a comparatively rapid release profile, where the xanthan gum-based system demonstrated sustained released behavior. In vivo findings revealed that the gum Arabic formulation significantly accelerated wound contraction, achieved near-complete epithelialization within 14 days, and significantly outperformed both xanthan gum formulation and untreated controls. Notably, both formulation enhanced healing responses relative to control. In conclusion, the developed PL-loaded nanogel formulation represented an efficacious strategy for regenerative burn wound management.
Development of a Paper-Based Origami Biosensor for the Detection of Sweat Glucose
Students: Jafrin Majumder Abony
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
This study presents a flexible, skin-attachable, noninvasive sweat-based colorimetric biosensor for glucose monitoring, offering a cost-effective, single-use alternative to traditional blood-based methods. Fabricated from Whatman filter paper and paraffin film, the sensor operates without an external power source through a passive enzymatic reaction. Glucose quantification relies on a cascade involving glucose oxidase, horseradish peroxidase, and 3,3′,5,5′-tetramethylbenzidine (TMB), producing a blue color whose intensity correlates with glucose concentration. This allows for easy qualitative assessment by the naked eye or quantitative analysis via a smartphone. Enzyme immobilization was enhanced using chitosan and chitosan-PBS composites, improving enzymatic stability and sensitivity. Validation with artificial sweat containing glucose concentrations from 0.055 to 1.5 mM demonstrated a consistent inverse relationship between R-channel intensity and glucose levels, supporting a linear regression model. Surface modification increased the R-channel response range from 60 to 106 units, with detection limits of approximately 0.0213 mM and 0.0321 mM for chitosan-only and dual PBSchitosan treatments, respectively. A dedicated smartphone application was developed for realtime colorimetric data processing, enabling precise remote monitoring. This integrated, powerfree platform advances personalized healthcare by facilitating continuous glucose monitoring in clinical and remote environments.
Development of an Antibacterial Drug-Loaded Gelatin-CMC Nanofiber Dressing Crosslinked with Citric Acid for Burn Wound Healing.
Students: Maharonnasa Al Tabassom
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Burn wound management remains challenging due to high infection risk, excessive exudate, and secondary tissue damage caused by adherent conventional dressings. This study aims to develop a biodegradable, electrospun nanofibrous dressing capable of providing sustained antimicrobial activity while promoting atraumatic wound healing. The scaffold was fabricated using natural polymers gelatin and carboxymethyl cellulose (CMC), crosslinked with citric acid to enhance structural stability and control degradation, and loaded with the broad-spectrum antibiotic tigecycline. The fabricated nanofibers exhibited a highly uniform, interconnected porous architecture closely resembling the native extracellular matrix, thereby providing an ideal microenvironment for cellular attachment and proliferation. Crosslinking significantly enhanced the mechanical integrity and enabled a tunable degradation profile aligned with the critical wound healing window (3-5 days). The hydrophilic nature of CMC conferred superior fluid absorption capacity and facilitated the formation of a hydrogel-like interface upon exudate contact, effectively minimizing tissue adhesion and secondary injury during dressing removal. In vitro evaluations indicated consistent antibacterial efficacy and hemocompatibility. Furthermore, in vivo studies demonstrated progressed wound healing characterized by organized collagen deposition, re-epithelialization, and accelerated wound closure compared to control groups. This study presents a multifunctional nanofibrous dressing that integrates structural support, moisture balance, and sustained drug delivery, offering a promising strategy for advanced burn wound care with reduced risk of infection and improved patient comfort.
Development of Chitosan-Carboxymethyl Cellulose Based Hemostatic Powder with Vasoconstriction Property for Rapid Hemostasis
Students: Alif Rudaba
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Uncontrolled bleeding is a major cause of death in trauma patients, making effective topical hemostatic agents essential. In this study, a novel hemostatic powder formulation was developed based on chitosan (CS) and carboxymethyl cellulose (CMC) to achieve rapid hemostasis with enhanced functionality of vasoconstriction. The formulation leverages the polyelectrolyte complexation between CS and CMC, which increases the swelling property of the powder that incorporates instant blood clotting efficiency, crosslinked with CaCl₂ to improve structural stability and biofunctionality. To achieve vasoconstriction property, an Adrenaline drug has been introduced with this formulation. This powder aims to address existing gaps in commercial products – particularly the lack of rapid vasoconstriction activity and insufficient swelling performance in local markets. Various formulations were prepared and optimized by altering CMC and CaCl2 concentration and physical state, and in vitro and in vivo tests were conducted to assess their effectiveness. The optimized formulation with and without vasoconstrictor drug, achieved a clotting time of 17.3 ± 2 sec and 9.1 ± 2 sec respectively in a mice liver laceration model and 25.5 ± 3 sec and 15.3 ± 3 sec respectively in a rat tail amputation model with BCI result of <5% in invitro test and demonstrated promising characteristics under ATR-FTIR and SEM characterization. In an invitro swelling test with SBF, optimised sample showed an increased water swelling property of 875.8% ±25% with compared to existing commercial product (Celox), which has swelling capacity of 432.5% ±30%.The results confirm the formulation’s potential for rapid hemostasis via enhanced swelling and local vasoconstriction property.
Development of Moxifloxacin loaded Sodium Alginate-based Nanoparticles with Enhanced Diffusibility for the Treatment of Ocular Infection and Inflammation.
Students: Omi Banik
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Conventional ocular therapies suffer from low bioavailability, rapid clearance, and poor penetration across ocular barriers. This research aimed to develop a novel drug delivery system using sodium alginate-based nanoparticles modified with several penetration enhancers like capric acid, citric acid, EDTA; to enhance drug diffusibility and retention in the eye. Moxifloxacin, a broad-spectrum antibiotic, was loaded into these nanoparticles synthesized via ionotropic gelation. TEM and DLS studies confirm variation in particle size and surface charge. The formulations were further evaluated through ATR FTIR, visual stability assessments, in vitro drug release studies, mucoadhesion, ex vivo Franz diffusion on excised goat eyes, in vivo pharmacokinetic profiling tests in rabbit eyes. The FTIR analysis indicated successful chemical interaction between the penetration enhancers and sodium alginate. Drug release studies revealed a sustained release pattern compared to commercial formulations. Mucoadhesion results showed improved adherence to ocular surfaces, suggesting potential for prolonged retention and reduced dosing frequency. Ex vivo permeation studies demonstrated increased transcorneal drug diffusion, while in vivo pharmacokinetic analysis revealed improved ocular bioavailability. This study offers a promising, non-invasive strategy for targeted treatment of ocular infections and inflammation.
Development of Polycaprolactone-Silk based Nanofibrous Mesh for Potential Hernia Repairment
Students: Mahema Akter Saj
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Hernia repair is a common surgical procedure; however, conventional non-biodegradable meshes are frequently associated with chronic inflammation and prolonged foreign body response. This study reports the development of a biodegradable hernia mesh based on an electrospun blend of polycaprolactone (PCL) and silk fibroin. PCL provides excellent mechanical strength and slow degradation, while silk fibroin enhances biocompatibility, hydrophilicity, and cellular interactions, enabling the fabrication of a nanofibrous scaffold that mimics the native extracellular matrix. Electrospinning parameters, including polymer concentration, applied voltage, flow rate, and tip-to-collector distance, were optimized to achieve uniform, bead-free fibers. A key finding of this study is that solution preparation time critically governs fiber morphology; prolonged stirring or storage leads to polymer degradation in formic acid, resulting in bead formation and non-uniform fibers, whereas freshly prepared solutions produce continuous and homogeneous structures. The optimized scaffold is expected to exhibit suitable mechanical integrity, porosity, and controlled degradation for effective tissue integration. Overall, this work highlights a simple yet crucial parameter for improving reproducibility and performance, demonstrating the potential of PCL-silk fibroin meshes as resorbable alternatives for hernia repair.
Epidemiologically-Informed Guided Attention for Oral Cavity Lesion Classification from Smartphone Imagery
Students: Samir Ahammad, Imdadul Haque Sourav
Supervisor: Dr. Taufiq Hasan, Professor
Abstract:
Oral cancer is a severe public health challenge in South Asia where the vast majority of cases are diagnosed at late stages due to limited screening infrastructure. While risk factors like betel quid chewing, smoking, and alcohol consumption are well documented population level predictors, current artificial intelligence screening tools treat images in isolation and ignore this vital clinical context. The primary objective of this project is to develop a multitask deep learning framework that accurately classifies smartphone captured oral cavity images into four categories: Healthy, Benign, Oral Potentially Malignant Disorders, and Oral Cavity Anomalies. The core methodology introduces an Epidemiological Prior Layer which encodes literature derived risk ratios as a learnable soft regularization mechanism. This allows the network to learn jointly with classification, shifting predictions toward clinically plausible outcomes based on patient habits. Furthermore, a Guided Attention Pooling mechanism uses expert annotated segmentation masks through Dice loss to ensure the network focuses strictly on relevant lesion regions. Dual classification heads simultaneously provide a four class diagnosis and a binary nonsuspicious or suspicious triage. Evaluated on 3000 images from 714 patients, the framework successfully achieved 71 percent four class accuracy, 82 percent binary screening accuracy, and 0.85 precision for malignant disorders. The learned epidemiological weights correctly identified betel quid as the highest risk factor, validating its consistency with regional profiles. Ultimately, this project offers a highly accessible and context aware screening application that could significantly improve early oral cancer detection rates in vulnerable populations, with prospective field validation currently underway in Dhaka Bangladesh.
Fitzpatrick-Guided Conditional GAN for Skin Tone Translation to Mitigate Racial Bias in Dermatological AI Systems
Students: Sharoare Hosan Emon
Supervisor: Dr. Taufiq Hasan, Professor
Abstract:
Modern artificial intelligence (AI) systems in dermatological disease detection have demonstrated high performance; however, their effectiveness is significantly limited across diverse demographic groups due to inherent dataset bias. Most traditional AI models are predominantly trained on datasets containing lighter skin tones, resulting in reduced accuracy and unreliable predictions for individuals with darker skin. This imbalance raises critical concerns regarding fairness, inclusivity, and real-world applicability in medical diagnostics. To address this issue, this research proposes a bias-aware image-to-image translation framework that generates diverse skin tone representations while preserving clinically relevant features. A specialized dataset was constructed based on the Fitzpatrick skin type scale, where original dermatological images were manually transformed into six distinct skin tone categories, ensuring controlled and balanced representation across demographic variations. A conditional Generative Adversarial Network (cGAN) was employed to learn the mapping between different skin tones. The model incorporates label conditioning through one-hot encoding, enabling targeted translation of images into specific skin tone categories. To ensure both visual realism and medical integrity, the training process integrates multiple loss functions, including adversarial loss, L1 reconstruction loss, perceptual loss using a pre-trained VGG19 network, and edge preservation loss using Sobel filters. These combined objectives allow the model to maintain fine-grained lesion characteristics while adapting skin tone variations. The experimental results demonstrate that the proposed approach can generate realistic and diagnostically consistent images across all Fitzpatrick skin types. By augmenting datasets with balanced skin tone variations, the model helps overcome the limitations of traditional AI systems that fail to generalize across demographic groups. This work contributes toward reducing racial and demographic bias in medical AI and supports the development of more equitable and robust disease detection systems.
Geometry-Informed Attention-Guided Unsupervised Domain Adaptation for Cross-Institutional Pediatric Pneumonia Classification from Chest Radiographs
Students: Iftakhar Hossain, Md. Imran Hasan
Supervisor: Dr. Taufiq Hasan, Professor
Abstract:
Pneumonia remains a leading cause of mortality among children under five globally, where timely and accurate radiological diagnosis is critical yet often constrained by limited access to trained clinicians. Deep learning models trained on institutional chest X-ray datasets have demonstrated diagnostic-grade performance in controlled settings; however, they suffer severe performance degradation under cross-institutional deployment due to distributional shifts arising from heterogeneous imaging equipment, acquisition protocols, and patient demographics, a fundamental barrier to clinical translation in low-resource environments. This thesis proposes an attention-guided unsupervised domain adaptation (UDA) framework for crossinstitutional pediatric chest X-ray classification that transfers diagnostic knowledge from a labeled source domain to an entirely unlabeled target domain. The framework integrates DenseNet-169 as a shared feature extractor augmented with a Convolutional Block Attention Module (CBAM) for domain-invariant semantic refinement. Adaptation is achieved through a unified multi-objective loss combining supervised source classification, entropy minimization, semantic consistency regularization, Conditional Domain Adversarial Network (CDANN) alignment with gradient reversal, and Maximum Mean Discrepancy (MMD) regularization. Transfer direction is principally selected via a comprehensive domain characterization analysis spanning PCA, t-SNE, convex hull and alpha shape containment, bounding hypersphere geometry, nearest neighbor distance distributions, MMD, and asymmetric KL divergence which establishes Sylhet as the more compact and distributionally regular domain, empirically justifying its role as source. Baseline evaluation reveals a pronounced asymmetric domain gap: a Kermany-trained model degrades to 56.65% accuracy (AUC: 66.48%) on Sylhet, while a Sylhet-trained model retains 84.78% (AUC: 92.40) on Kermany. The proposed framework substantially mitigates this gap in both directions: Sylhet→Kermany adaptation achieves 94.71% accuracy (AUC: 98.97), approaching the within-domain upper bound of 95.35%, while Kermany→Sylhet adaptation improves from 56.65% to 70.70% (AUC: 75.06). These results demonstrate that principled domain selection combined with attention-guided feature refinement, conditional adversarial alignment, and kernel-based distribution matching yields a robust and clinically relevant UDA strategy for diagnostic AI deployment in annotation-scarce environments.
Ionotropically Gelled Mucin-Carboxymethyl Cellulose Nanoparticles for Ocular Surface Retention in Dry Eye Disease
Students: Tasfia Tabassum
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Dry eye disease is a common ocular disorder which results from tear film instability and inadequate lubrication of the ocular surface. Although conventional artificial tear formulations, such as carboxymethyl cellulose, provide temporary relief but are limited by rapid clearance from the ocular surface and their limited therapeutic efficacy. This study explored the potential of enhancing mucoadhesion through the incorporation of mucin into a carboxymethyl cellulose matrix and the development of a nanoparticulate delivery system using the ionotropic gelation method. This study aimed to investigate whether incorporating mucin into a carboxymethyl cellulose based formulation and converting it into a nanoparticulate system could enhance mucoadhesion and improve retention. Nanoparticle formulations were prepared and assessed for mucoadhesive properties using ex vivo methods. A comparative evaluation between standard carboxymethyl cellulose and the mucin incorporated formulation showed a slight increase in mucoadhesion. In parallel, in vivo animal studies have been initiated to assess ocular surface healing and retention characteristics of the developed formulation. Preliminary observations suggest a trend toward improved therapeutic response in the mucin containing formulation, although comprehensive analysis is still ongoing. Overall, the current findings indicate that while ex vivo mucoadhesion improvement is limited, further investigation through in vivo studies and formulation optimization is necessary to fully understand the potential benefits of mucin incorporation in dry eye therapy.
Laser Ablation of Breast Tumours: a computational study and analysis on different parameters affecting ablation
Students: Mubtasim Fuad Sami
Supervisor: Dr. Jahid Ferdous, Associate Professor
Abstract:
Laser-induced thermal therapy is a minimally invasive approach for treating breast tumors, directing laser energy into tissue where it converts into heat to destroy cancer cells. Driven by the primary goal of deducing optimized treatment protocols, this research develops a threedimensional computational model utilizing patient-specific geometries of breast tissue containing tumors. The simulation employs a two-domain framework consisting of biological tissue and a cylindrical optical diffuser. While various techniques model light absorption, this study specifically uses a PDE-based diffusion approximation to compute the fluence rate. Highly suited for the intense light scattering characteristic of breast tissue, this method is coupled with the bioheat transfer equation to map temperature distribution over time. Thermal damage is subsequently quantified using Arrhenius kinetics to estimate tissue necrosis. The core objective is to execute these simulations across breast geometries from different patients, enabling robust comparative studies to evaluate treatment efficacy. The exemplary two domain model assesses how altering key variables specifically laser power inputs (7.4 W and 15 W) and spatial geometry impacts the temperature gradient during ablation. By analyzing spatial distributions of fluence rate and temperature alongside the temporal evolution of thermal damage, results demonstrate strong tissue sensitivity to these parameters. While higher energy inputs yield significantly greater ablation volumes, comparative findings indicate that for larger tumors, the overall necrotic fraction remains comparatively low. This highlights the critical influence of patient-specific anatomy and tumor size on treatment efficiency. To further enhance predictive accuracy, future work will integrate temperature-dependent properties across a broader patient cohort. Ultimately, this patient-specific computational framework provides a valuable clinical tool for representing realistic conditions, actively supporting the deduction of safer, personalized thermal ablation strategies.
Modeling and Simulation of PEG-PLA Polymeric Micelle Self-Assembly Using Coarse-Grained Molecular Dynamics
Students: Dil Marufa Tarannum Megha
Supervisor: Dr. Jahid Ferdous, Associate Professor
Abstract:
This study explores the self-assembly behavior of polymeric micelles using coarse-grained molecular dynamics (CG-MD) simulations. The main objective is to develop a basic understanding of micelle formation without the influence of drug molecules or organic solvents. Simulations were performed using the GROMACS software with the MARTINI coarse-grained force field. PEG-PLA block copolymers were modeled in an aqueous environment, and equilibrium was achieved through NVT and NPT simulations. The resulting trajectory and structural data were analyzed to assess system stability and early-stage aggregation behavior. The results show successful system equilibration along with initial signs of polymer aggregation, indicating the onset of micelle formation. Studying the system without paclitaxel and ethanol allows for a clearer understanding of the intrinsic interactions between polymer chains. This work serves as a foundation for future studies that will incorporate drug loading, solvent effects, and temperature variations to better understand micelle stability and drug delivery performance. The findings are expected to support the development of efficient polymeric micelle-based drug delivery systems.
Modeling Targeted Magnetic Hyperthermia in Endometriosis
Students: Jarin Tasnim Joyee, Sadia Afrin
Supervisor: Dr. Jahid Ferdous, Associate Professor
Abstract:
Endometriosis is a chronic gynecological disorder characterized by the growth of endometrial-like tissue outside the uterus, leading to severe pelvic pain and infertility in millions of women worldwide. Magnetic nanoparticle-mediated hyperthermia has recently emerged as a promising non-invasive therapeutic strategy; however, its effectiveness depends on efficient nanoparticle delivery, targeting, and heat generation. Inspired by recent advances in systemically delivered magnetic hyperthermia, this study develops a comprehensive computational model to investigate the transport, accumulation, and thermal performance of both non-targeted and KDR-targeted magnetic nanoparticles in endometriotic lesions. Iron-based nanofluids, both non-targeted and KDR-targeted, are modeled following intravenous administration, incorporating systemic circulation, extravasation, and selective binding to endometrial receptors. A coupled multiphysics model is developed in COMSOL to solve heat transfer, nanoparticle kinetics, and bioheat interactions within the tumor microenvironment. Thermal damage is quantified using the Arrhenius kinetic model, along with a threshold-based analysis of tumor regions exceeding 48 °C. Results demonstrate that KDR-targeted nanoparticles significantly enhance intralesional accumulation, leading to higher temperature elevation and greater thermal damage compared to non-targeted counterparts. The study highlights the potential of targeted magnetic hyperthermia as an effective treatment strategy for endometriosis and provides a predictive tool for optimizing nanoparticle design and treatment parameters for improved therapeutic outcomes.
Molecular Dynamics Simulation of Adhesion and Cohesion Behavior of Drug-Coated Balloons: A Computational Approach
Students: Rudmila Nizam, Sunaina Rahman Adisha
Supervisor: Dr. Jahid Ferdous, Associate Professor
Abstract:
Drug-coated balloons are widely used in angioplasty for treating peripheral arterial disease by enabling localized drug delivery to the vessel wall. However, significant drug loss during delivery, inflation, and retraction reduces transfer efficiency and limits therapeutic effectiveness. The main objective of this analysis is to investigate the adhesion between the coating layer and balloon surface, as well as the cohesion within the coating material, to improve drug retention performance. Molecular dynamics (MD) simulations are employed to analyze interfacial behavior at the molecular level. A simplified three-layer model consisting of arterial tissue, drug-excipient coating, and balloon material is developed using CHARMM-GUI, Packmol, Material Studio, and VMD, with simulations performed in LAMMPS. Various excipients and drug-excipient ratios are considered to evaluate their influence on interfacial stability and mechanical properties. The system is energy minimized, equilibrated under controlled temperature, and subjected to pulling forces to simulate mechanical separation during balloon withdrawal. Stress-strain responses and structural changes are analyzed. Preliminary results indicate interfacial separation, localized void formation, and progressive coating deformation under mechanical loading. These findings provide insight into factors affecting coating integrity and drug retention. This study contributes to the optimization of coating materials and design, potentially enhancing the efficiency and reliability of DCBs in clinical angioplasty applications.
Multimodal Waveform-Based Prediction of Intraoperative Vasopressor Requirement
Students: Jesia Briti, Mir Mashrafi Ahasan, Moshiur Rahman
Supervisor: Dr. Jahid Ferdous, Associate Professor
Abstract:
Intraoperative hypotension is independently associated with acute kidney injury, myocardial injury, and increased perioperative mortality, yet clinical management remains reactive. Vasopressors are administered only after hypotension is established, by which point organ hypoperfusion may already be underway. This study investigates early prediction of intraoperative vasopressor requirement using deep learning on multimodal physiological waveforms, aiming to enable proactive haemodynamic intervention. Arterial blood pressure, photoplethysmography, and electrocardiography waveforms were extracted from VitalDB and preprocessed through physiologically-grounded filtering, interpolation, resampling, and per-segment Z-score normalisation, with strict patient-level data isolation enforced throughout. Performance was assessed using AUROC, AUPRC, and F1-score with 95% bootstrapped confidence intervals, alongside Brier score for calibration. A systematic progression from convolutional baselines through recurrent, hybrid, and attentionbased architectures revealed that CNN-based feature extraction paired with attention consistently outperforms pure transformer approaches on raw waveform inputs. A patch-based transformer evaluated across four prediction horizons showed the trimodal combination achieving AUROC of 0.937 at five minutes; AUROC of 0.857 at ten minutes; and AUROC of 0.833 at fifteen minutes – with Brier scores of 0.127, 0.162, and 0.214 respectively, indicating well-maintained calibration across horizons. ABP was identified as the dominant modality at shorter horizons, with performance degrading expectedly at longer lead times. These findings establish a rigorous architectural and modality selection framework for future real-time intraoperative decision support systems.
Nanoemulgel-Based Topical Delivery of Wheat-Derived Antimicrobial Peptides for Antibiotic-Resistant Wound Healing
Students: Tasmia Tabassum Ananna
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
The rapid emergence of antimicrobial resistance has intensified the need for bioactive, biocompatible, and sustainable alternatives for wound management. Wheat-derived antimicrobial peptides (AMPs) are promising candidates for treating infected wounds due to their broad-spectrum antibacterial activity and lower likelihood of inducing resistance. However, their therapeutic application remains constrained by poor stability, susceptibility to degradation, and inadequate penetration at the target site. Recent studies suggest that nanoemulsions (NEs) can protect labile bioactive compounds from degradation and thereby enhance their stability. Therefore, the study developed a nanoemulgel-based delivery system for wheat-derived AMPs. A few NE formulations with varying peptide contents were designed first to encapsulate the AMPs, and the optimized formulation was subsequently incorporated into 1.5% Carbopol gel base to enable convenient topical application. The formulations were evaluated through antibacterial assays, physicochemical characterization and animal trials to assess release behavior, structural integrity, stability, antimicrobial activity, and wound healing performance. The results demonstrated successful AMP release from the NE platform, while TEM analysis confirmed peptide protection within nanodroplets. The optimized nanoemulgel exhibited satisfactory physicochemical stability and effective antibacterial activity against multiple bacteria, including strains resistant to many conventional antibiotics. In vivo animal studies further confirmed excellent wound healing efficacy with nearly complete wound closure. Collectively, these findings indicate that incorporation of wheat-derived AMPs into a nanoscale platform is an effective strategy to overcome the major barriers associated with free peptide administration. This study establishes a novel topical nanodelivery strategy for natural AMPs as an alternative therapeutic platform for infected and antibiotic-resistant wound management.
Non-contact Estimation of Pulse and Blood Pressure from Facial Videos
Students: Tahmidur Rahman, Sadman Sakib Himel
Supervisor: Dr. Taufiq Hasan, Professor
Abstract:
Continuous monitoring of heart rate and blood pressure typically requires physical contact with the patient, limiting its use in remote or low-resource settings. Remote photoplethysmography (rPPG) offers a camera-based alternative, but existing models perform poorly on underrepresented demographic groups. This study develops and evaluates an rPPG pipeline tailored to demographic groups underrepresented in existing datasets, with a focus on South-Asian dark-skinned individuals. Building upon the CAN2DShare model proposed by Dasari et al., we developed a new dataset comprising 107 subjects (target:120), including synchronised facial video recordings, continuous heart rate signals from a pulse oximeter, and averaged blood pressure measurements as ground truth. The pretrained model is fine-tuned using this dataset to adapt it to demographic-specific characteristics. Additionally, spatial calibration is performed by estimating shoulder width from video frames, while subject-specific parameters such as height and weight are incorporated to approximate vascular path length. These features are further utilised with the Moens-Korteweg equation to estimate blood pressure. Preliminary results show diastolic BP estimation achieving an MAE of 3.23 mmHg and a Pearson correlation of 0.880, and systolic BP at an MAE of 7.69 mmHg, correlation of 0.780, with 97.1% of diastolic predictions falling within ±10 mmHg of ground truth. Heart rate estimation accuracy is 93.15% with a MAE of 6.02 BPM. These results confirm that the V-BPE biophysical model paired with demographic-aware rPPG yields clinically acceptable diastolic BP estimates, though systolic overestimation bias indicates that further calibration of anthropometric parameters is needed before deployment.
Optimizing Convection Enhanced Delivery for Patient-Specific Drug Delivery in Brain Cancer: An In Silico Approach
Students: Mohammad Samsul Arefin Safi, Mashiat Subah Abonty
Supervisor: Dr. Jahid Ferdous, Associate Professor
Abstract:
Brain cancer treatment remains challenging due to the limited effectiveness of conventional drug delivery methods in overcoming the blood-brain barrier. Convection-enhanced delivery (CED) offers an emerging strategy by enabling direct infusion of therapeutic agents into brain tissue; nevertheless, the anisotropic characteristics of the brain that impact drug distribution limit the effectiveness of treatment. This study presents an in silico approach to investigate drug transport and therapeutic performance during CED. A three-dimensional finite element model reconstructed from Magnetic Resonance Imaging (MRI) data to simulate drug deposition in brain tumors. Brain tissue anisotropy was incorporated using Diffusion Tensor Imaging (DTI) obtained from MRI data to characterize directional diffusion behavior. Tumor volume and the location of the infusion site were assessed to reflect various clinical situations. Simulation results demonstrate that multi-probe catheter configurations provide improved spatial control of drug distribution compared with single-probe systems. Drug-tissue binding kinetics were also modeled to analyze temporal variations in free and bound drug concentrations. The proposed computational methodology assists the optimization of CED techniques for improved therapeutic outcomes for brain cancer and offers insights into the spatiotemporal distribution of drugs within lesions.
Plantar pressure assessment of diabetic patients for early diagnosis of diabetic foot ulcer
Students: Zurafa Najiat, Israt Jahan
Supervisor: Dr. Jahid Ferdous, Associate Professor
Abstract:
Diabetic Foot Ulcer (DFU) is one of the most severe complications of diabetes mellitus and a major cause of lower-limb amputation worldwide. In developing countries like Bangladesh, the risk is further amplified by limited awareness of foot care, socioeconomic constraints, and cultural footwear practices. Although pressure-offloading footwear and insoles are widely recommended for prevention, accessible and cost-effective solutions suitable for local populations remain scarce. This thesis focuses on the preliminary development of a DFU-preventive insole tailored to the socioeconomic and cultural context of Bangladesh. At this stage, the study emphasizes the collection and analysis of plantar pressure data using a pressure measurement system. Data are being collected from both diabetic and healthy individuals to enable a comparative evaluation of pressure distribution patterns. The primary objective is to identify high-pressure regions on the plantar surface that may contribute to the early onset of foot ulcers. The research is currently centered on biomechanical pressure analysis as a preventive and early detection approach. The findings will help establish pressure distribution characteristics and highlight potential risk zones. This dataset will serve as a foundation for designing a customized insole capable of effectively redistributing plantar pressure. In future phases, additional clinical factors, particularly diabetic peripheral neuropathy, will be incorporated to enhance the design. By integrating biomechanical insights with clinical considerations, this study aims to develop a practical, affordable, and culturally appropriate DFU-preventive solution for diabetic populations in Bangladesh
Stain-Invariant HER2 Prognosis from IHC to H&E via Cross-Modality Knowledge Distillation
Students: Fariha Hasan Chowdhury, Nibir Saha, Sumaiya Islam Shotota
Supervisor: Dr. Taufiq Hasan, Professor
Abstract:
Breast cancer is the most common cancer affecting women with roughly 15-20% of cases involve Human Epidermal Growth Factor Receptor 2 (HER2) amplification, making the disease behave more aggressively. This study evaluates a deep learning approach for achieving more consistent classification via IHC or Hematoxylin and Eosin(H&E) images. Our study utilized the BCI dataset (Liu, Shengjie et al.), which includes 4,870 paired H&E and HER2 images along with additional samples collected from Labaid Cancer Hospital for generalizability. The proposed methodology is based on a three-phase knowledge distillation framework called HER2-SINet for automated HER2 scoring (0, 1+, 2+, 3+) from histopathology images. The objective was to develop a single model capable of handling different staining modalities while predicting HER2 expression levels across four classes. A teacher model is first trained on IHCstained images where HER2 expression is visible, then its knowledge is transferred to a student model via a multi-component distillation loss combining soft KL divergence, feature alignment, class-weighted cross-entropy, and ordinal distance penalization which subsequently fine-tuned to produce a single stain-invariant classifier that accepts either IHC or H&E input without requiring stain-type annotation at inference. The results demonstrate that transferring knowledge from IHC to H&E is effective for improving cross-modality understanding and the GradCAM++ visualizations provide interpretable evidence that the model attends to clinically relevant tissue features in both modalities. This approach highlights the potential of distillation-based learning to bridge the gap between staining techniques and supports the development of flexible, modality-agnostic diagnostic systems.
Topology-Based Structural Encoding for Interpretable Knee Osteoarthritis Grading from Radiographs
Students: Shafqat Alam, Galib Anjum Talukder Mahi
Supervisor: Dr. Taufiq Hasan, Professor
Abstract:
Automated assessment of knee osteoarthritis (KOA) severity from radiographs remains limited by image-only models that fail to explicitly capture joint-level structural and biomechanical degeneration. We propose a topology-based structural encoding (TBSE) framework that integrates anatomically grounded joint geometry into deep learning for improved Kellgren-Lawrence (KL) grading. A landmark detection network localizes 148 anatomical landmarks with subpixel-level accuracy (mean error: 1.38 pixels), enabling reconstruction of joint structure and estimation of joint space width (JSW). From these landmarks, the proposed representation captures clinically relevant characteristics, including joint space narrowing, surface congruency, and medial-lateral asymmetry. The structural encoding is incorporated into a convolutional backbone via an adaptive gating mechanism that leverages bilateral differences to guide prediction. This allows the network to combine visual features with topology-aware cues while preserving ordinal relationships between KL grades. Trained on the Osteoarthritis Initiative dataset, the model explicitly encodes disease-relevant structure rather than relying solely on appearance. The model achieves a quadratic weighted kappa of 0.8610 and an overall accuracy of 0.7195, outperforming image-only baselines and demonstrating improved ordinal consistency (mean absolute error: 0.3180). The structural features provide complementary information that reduces ambiguity between adjacent KL grades. Landmark-derived JSW measurements show high geometric fidelity, with mean errors of 0.4306 mm, supporting reliable structural characterization. By embedding joint topology and biomechanics into deep learning, the proposed approach enables more interpretable and clinically consistent KOA severity assessment. Evaluation on independent cohorts is ongoing to further assess generalizability and clinical utility.
Wheat Extracted Antimicrobial Peptides Loaded Hydrogel for Antibiotic Resistant Wound Healing
Students: Kazi Fariha Farid
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
The increasing prevalence of antibiotic-resistant wound infections necessitates the development of alternative therapeutic strategies. This study aimed to develop an antimicrobial peptide (AMP) loaded biopolymeric hydrogel for effective management of antibiotic-resistant wound pathogens. Antimicrobial peptides were extracted from wheat using an appropriate extraction buffer and subsequently incorporated into a sodium alginate-gelatin hydrogel matrix to fabricate a bioactive wound dressing scaffold. Dopamine was introduced into the hydrogel system to enhance peptide stability, facilitate functionalization, and improve intermolecular interactions within the polymeric network. The developed hydrogel was evaluated for its antibacterial efficacy against antibiotic-resistant pathogenic Staphylococcus aureus, a major causative agent of chronic wound infections and biofilm formation. The AMP-loaded hydrogel demonstrated significant antimicrobial activity, suggesting sustained peptide release and effective bacterial growth inhibition. The presence of dopamine is expected to contribute to improved peptide immobilization and enhanced structural integrity of the hydrogel. Overall, the formulated AMPbased hydrogel presents a promising biomaterial platform for combating antibioticresistant wound infections and promoting improved wound healing outcomes, highlighting its potential application as an alternative to conventional antibioticbased therapies.