Influence of Anthropometric Factors on Gait Kinematics
Students: Farjana Rahman, Md. Redwan Hossain, Abrar Ahmed
Supervisor: Dr. Jahid Ferdous, Associate Professor
Abstract:
Gait is a biomechanical process influenced by anthropometric factors such as Body Mass Index (BMI) and Body Roundness Index (BRI), which impact key gait parameters, including cadence, velocity, stride length, and gait complexity. This study analyzed the gait patterns of 109 participants using video-based motion tracking and force plate measurements, with statistical evaluations performed to assess the effects of body composition on walking dynamics. The results revealed that individuals with higher BMI and BRI exhibited lower cadence and velocity, prolonged step time, and increased gait complexity, indicating biomechanical adaptations likely due to stability and energy efficiency constraints. While step length showed a decreasing trend with increasing BMI, stride length demonstrated an unexpected increase in obese individuals, suggesting possible compensatory mechanisms. Additionally, frequency dispersion and gait regularity index values varied across BMI and BRI categories, highlighting greater movement irregularities in individuals with higher body mass and roundness. These findings underscore the importance of incorporating both BMI and BRI in gait assessments, as BRI captures body shape variations that BMI alone does not. The study has implications for clinical gait analysis, rehabilitation, and mobility research, emphasizing the need for personalized movement assessments based on body composition. Future research should integrate kinetic analysis and muscle activity tracking to provide a more comprehensive understanding of gait adaptations in individuals with varying anthropometric characteristics.
Mitigating Demographic Bias in Skin Lesion Classification Using Deep Learning Models: A GAN-Based Approach
Students: Mobaswir Al Farabi, Nazmus Sadat
Supervisor: Dr. Taufiq Hasan, Professor
Abstract:
Machine learning models for skin lesion classification often exhibit biases due to the underrepresentation of darker skin tones in dermatology datasets. This disproportion leads to lower diagnostic accuracy for patients with darker skin, raising concerns about fairness and generalizability in AI-driven healthcare. To address this issue, we propose a generative approach to mitigate skin tone bias by synthesizing diverse skin lesion images while preserving the characteristics of the lesion. Our methodology leverages Pix2Pix and CycleGAN to generate skin lesion images across different Fitzpatrick skin tones. Pix2Pix is employed for paired image-to- image translation using semantic maps, enabling controlled background skin tone modifications while maintaining lesion integrity. Meanwhile, CycleGAN is utilized for unpaired domain adaptation, transforming real clinical images from datasets such as DermNet and Fitzpatrick17k into multiple skin tones without requiring explicit pixel-wise correspondences. To ensure accurate tone translation, a skin tone classifier is integrated into the training pipeline, guiding the GANs through an additional loss term that reinforces proper skin tone generation. Experimental results demonstrate that augmenting lesion classification models with these synthetic images significantly reduces performance disparities across skin tones. By bridging the gap in data representation, our approach enhances fairness in dermatological AI, promoting more equitable diagnostic outcomes for all skin types.
AI-Driven Anthropometry and Diagnosis: Leveraging PiFuHD and LLMs for Tropical Medicine
Students: S. M. Sakeef Sani, Md. Shaown
Supervisor: Dr. Taufiq Hasan, Professor
Abstract:
Malnutrition and complex differential diagnoses remain pressing challenges in global health, particularly in low-resource settings with limited healthcare infrastructure and skilled personnel. This thesis presents an end-to-end AI-driven solution integrat- ing smartphone-based anthropometric measurements with advanced clinical decision- making tools to address these issues. We introduce a novel system that leverages a multilevel Pixel-Aligned Implicit Function (PIFuHD) framework to generate high- resolution 3D meshes from single smartphone-captured images, enabling precise ex- traction of anthropometric parameters—such as chest, waist, thigh, and approximated weight—via cross-sectional analysis. Complementing this, we develop mLabLLM, a fine-tuned LLaMA 3.2 3B model optimized with Low-Rank Adaptation (LoRA) and pruning, tailored for differential diagnosis in tropical medicine. Trained on a curated dataset of tropical diseases prevalent in South Asia (e.g., Dengue, malaria, chikun- gunya), mLabLLM employs a probabilistic Bayesian framework to dynamically rank diagnoses based on symptom-disease frequencies, achieving an 82.8% Top-3 accu- racy—outperforming baselines like Phi-3-128k (75.1%) and LLaMA 3.2 3B (72.4%). The anthropometric system enhances malnutrition screening by offering a scalable, low-cost alternative to conventional methods, which often rely on specialized tools and expertise, while mLabLLM supports clinicians in resource-constrained environ- ments by providing efficient, accurate diagnostic insights. Ethical considerations, in- cluding data privacy and equitable access, are prioritized throughout to ensure trust and compliance with global standards. Together, these innovations demonstrate the transformative potential of AI in improving health assessments and outcomes in un- derserved communities, combining advanced computer vision, machine learning, and language modeling to deliver a practical, accessible solution for early detection.
Carbodiimide Crosslinked Silk Fibroin with Unoxidized Tannic Acid/Gelatin-based Bioadhesive for Enhanced Functional Properties
Student: Chaity Chakraborty
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Tissue sealants are materials used in surgical and medical applications to close wounds, prevent leakage of fluids, and promote healing. They are commonly used when sutures or staples are insufficient or impractical. Tissue adhesives play a critical role in modern medicine by providing efficient wound closure, hemostasis, and tissue sealing. However, existing adhesives often suffer from limitations such as cytotoxicity, weak adhesion, and high costs. To address these challenges, a novel bioadhesive was developed using gelatin (GA), silk fibroin (SF), and tannic acid (TA), with SF modified through carbodiimide coupling (EDC/NHS) followed by dopamine (DOPA) functionalization. The adhesion strength (311.02±30.2 kPa for dry skin and 142.9943±25.3 kPa for wet skin) and toughness (~1567 kJ/m3 ) were significantly higher than previously reported tannic acid-gelatin based bioadhesives. Fourier-transform infrared (FTIR) spectroscopy confirmed the successful modification of SF and integration of TA and SF, highlighting hydrogen bond formation and cohesive interactions. The adhesive demonstrated prolonged underwater stability, maintaining adhesion for over ten days, and effectively sealed tissue wounds in mice skin model as well as could also join internal organs. Both the samples PG-TA-SF-EDC/NHS-DOPA and PG- TA-SF-DOPA demonstrated significant antibacterial efficacy against gram-positive (S. aureus) and gram-negative (E. coli) pathogens throughout the antibacterial investigation. The prepared samples also showed excellent hemostatic properties and blood clotting abilities. PG-TA-SF- EDC/NHS-DOPA was able to stop bleeding in mice liver in 15.25 seconds. The cross-linking among 20% (w/v) gelatin, 8% (w/v) SF modified with EDC/NHS and DOPA, and 20% (w/v) tannic acid was performed at room temperature facilitated by centrifugation process. Hydrochloric acid was used as an acidic medium to create an unoxidized environment. The innovation resides in both the improved adhesion strength and toughness, as well as the straightforward single-step production procedure of the adhesive. This study results in tough and antibacterial adhesives, with adhesion properties that can be fine-tuned by adjusting the pH. Designed for topical application, it shows promise for wound closure and could also serve as a multifunctional tissue adhesive, facilitating regulated medication administration at wound locations and many biomedical applications.
A Molecular Dynamics Study of Dextran-Chitosan Copolymers for Enhanced Aldehyde-Mediated Cohesion
Students: Nishat Tasnim, Tanjila Hossain Tamanna
Supervisor: Dr. Jahid Ferdous, Associate Professor
Abstract:
Bio-adhesive polymers are novel biomaterials of significant promise for wound healing and tissue engineering. Molecular dynamics (MD) simulations have been employed in this study to investigate the cohesive behavior, solubility, diffusion, and overall trend of the Chitosan Dextran-Aldehyde copolymer. Maximizing the distribution of reactive aldehyde groups was pursued to maximize molecular interactions in a consideration of biocompatibility. Major analyses included molecular interactions, hydrogen bonding, radial distribution functions (RDF), cohesive energy density (CED), and solubility parameters. Additional studies on adhesive properties are being conducted with a bulk simulation model with a protein in the vicinity for future studies. The present work provides insights into the rationale design of bio- adhesive polymers and their use in soft tissue adhesion, propelling biomaterial-based surgical sealants.
Development of Artificial Atherosclerotic Plaques to Better Understand the Endovascular Drug Delivery System
Students: Ayantika Das, Samiyee Islam
Supervisor: Dr. Jahid Ferdous, Associate Professor
Abstract:
Atherosclerosis is a progressive cardiovascular disease characterized by the accumulation of lipid-rich plaques within arterial walls, leading to reduced blood flow and an increased risk of severe cardiovascular events. Drug-Coated Balloon (DCB) therapy has emerged as an effective treatment modality, delivering antiproliferative drugs directly to the lesion site to inhibit restenosis. However, preclinical studies primarily rely on bovine artery models, which fail to accurately replicate the mechanical and physicochemical properties of human atherosclerotic plaques. Consequently, drug delivery assessments based on these models often lack clinical relevance, necessitating the development of more representative artificial plaque models. In this study, we developed synthetic atherosclerotic plaques designed to closely mimic the structural, mechanical, and biochemical characteristics of native plaques. The artificial plaques were fabricated using biomimetic materials and subjected to extensive mechanical testing, to compare their behavior with that of human atherosclerotic tissue. Additionally, physicochemical characterization techniques were employed to assess surface morphology, porosity, and composition. Following plaque characterization, a drug delivery study was conducted using DCBs to evaluate drug penetration, retention, and release kinetics. The results demonstrated that the artificial plaques provided a more reliable and reproducible platform for studying drug absorption and retention, addressing the limitations of existing ex vivo models. The improved plaque model enhances our understanding of drug transport mechanisms in atherosclerotic lesions and could serve as a valuable tool for optimizing DCB formulations and refining treatment strategies. This research contributes to the development of more clinically relevant in vitro models for drug delivery studies, ultimately aiding in the advancement of atherosclerosis treatment and improving patient outcomes
Deep Learning Models for Tuberculosis Detection from Chest X-rays
Students: Marshal Ashif Shawkat, Moidul Hassan
Supervisor: Dr. Taufiq Hasan, Professor
Abstract:
Tuberculosis (TB) remains one of the most significant global health challenges, particularly in resource-constrained settings where access to expert radiologists and advanced diagnostic tools is limited. Chest X-ray (CXR) imaging is a widely used preliminary diagnostic tool for TB. Recently, computer-aided tuberculosis diagnosis (CTD) using deep learning has shown promise, but current techniques have poor interpretability and generalisation ability. This thesis addresses the critical need for interpretable and generalizable TB detection systems by leveraging state-of-the-art deep learning techniques. We propose a comprehensive framework that uses knowledge distillation to enhance the localisation of TB from CXR images. Our results show that the proposed framework not only improves diagnostic accuracy but also provides interpretable visualisations of TB lesions, aiding radiologists in clinical decision-making. Besides, the incorporation of knowledge distillation enhances the model’s generalisation ability, achieving higher AUC scores across diverse datasets.
Surfactant Modified Sodium Alginate/Mucin Nanoparticles for Moxifloxacin Delivery
Student: Tahia Ifreet
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Novel surfactant-modified alginate-mucin nanoparticles were synthesized via the ionotropic gelation method, employing Ca2+ ions as crosslinkers for targeted eye infection treatment. The nanoparticles demonstrated a uniform spherical morphology and high stability, as confirmed by TEM imaging and zeta potential analysis. Additionally, they exhibited significant mucoadhesive properties. The optimal surfactant concentration was determined to be 0.1%, based on TEM and zeta potential assessments, and this formulation was selected as the delivery vehicle. The final nano formulation displayed shear- thinning behavior, making it well-suited for ocular applications. The nanoparticles were non hemolytic and did not exhibit any signs of corneal edema or cell damage in vivo and ex vivo. Moxifloxacin (MFX) was successfully encapsulated into the optimized nanoparticles, achieving an encapsulation efficiency of 98%. The MFX-loaded nanoparticles exhibited a Fickian release pattern in the appropriate release medium. Furthermore, different surfactant-modified nanoparticles displayed varied controlled release profiles, highlighting their potential for customizable drug delivery applications.
A Gauze Dressing Incorporated with Photosensitizer-Loaded Hydrogel for Combating Antibiotic-Resistant Bacteria
Student: Muzeza Jannat Shoshee
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
The rapid emergence of antibiotic-resistant bacteria poses a critical challenge to global health, which calls for creative treatment approaches. By combining cellulosic cotton gauze with photosensitizer-loaded gelatin-based hydrogels presents a novel wound dressing with the goal of improving moisture retention at the wound site and antibacterial qualities for therapeutic applications. The dressing uses photosensitizers including methylene blue and curcumin, which produce reactive oxygen species (ROS) for effective bacterial eradication. By facilitating hydrogel integration, this innovative dressing addresses the major drawbacks of conventional gauze—such as poor moisture retention and adherence. Thus, it not only maintains a moist wound environment but also prevents reinjury during dressing changes. Comprehensive evaluations, including swelling behavior, degradation, FTIR analysis, and SEM characterization, were performed to optimize material properties. To optimize the effectiveness of the drug at the wound site, drug loading capacity and drug release were also assessed. Antibacterial efficacy was validated through disk diffusion, series dilution and bacterial penetration assays. Initial results demonstrate significant antibacterial activity against multidrug-resistant pathogens, effective release of photosensitizers, and enhanced moisture retention. Furthermore, in vivo studies were conducted to validate clinical efficacy and to explore the underlying biological mechanisms. The findings from in vivo studies indicate a significant advancement in the treatment of wound infections. Therefore, this active dressing represents a novel and effective approach in combating wound infections, particularly addressing the challenges posed by the growing global crisis of antibiotic resistance.
An Antifibrinolytic Drug Loaded Gelatin/Alginate-Based Foam for Dual Hemostatic Efficacy with Enhanced Wound Healing
Student: Anisha Fairooz Borsha
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Excessive bleeding, whether resulting from surgery, trauma, or postpartum hemorrhage—presents a significant clinical challenge, particularly in patients with underlying coagulopathies. In this study, a novel gelatin-alginate based hemostatic foam was developed, employing physical crosslinking with CaCl2 and loading an antifibrinolytic drug aminocaproic acid to achieve dual hemostatic efficacy and enhanced wound healing. Functional properties, including crystallinity, pore size, swelling capacity, and compressive strength, were systematically evaluated across varying gelatin concentrations. An optimized polymer matrix was subsequently loaded with 0.5% of the drug, yielding the sample A3G3D, which demonstrated not only effective primary hemostasis through rapid blood absorption—as confirmed by in vitro whole blood clotting tests (BCI value: 4.08 ± 1.61%) and in vivo evaluations using a mice laceration model (bleeding time: 18.27 ± 1.63 s; blood loss: 63.25 ± 0.61 mg) and a rat tail amputation model (bleeding time: 20.05 s; blood loss: 60 mg)—but also superior secondary hemostatic efficacy, as evidenced by significantly lower PT (12.54 ± 0.32 s) and aPTT (25.09 ± 0.08 s) values compared to both blank controls and commercial product. Moreover, the foam exhibited high hemocompatibility (<2% hemolysis ratio), cellular adhesion capacity, favorable in vivo biodegradation, and promoted angiogenesis, thereby supporting its biocompatibility and wound healing potential. Overall, this hemostatic foam holds significant promise for surgical applications and for managing bleeding in patients with coagulopathies, offering a unique and effective alternative to existing solutions.
Carbomer-Based Essential Oil Nanoemulgel as Topical Formulations for Treating Infectious Diseases
Student: Srija Sarker Anannya
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Ciprofloxacin (Cp) loaded nanoemulsions (NEs) and nanoemulgels (NEGs) were formulated as potential topical application for treatment of infectious diseases. Lipophilic drugs in most cases show low bioavailability due to poor solubility and reduced absorption. Oil in water nanoemulsion can encapsulate these lipophilic drugs but the liquid form is difficult to apply and also have poor skin retention. So, nanoemulgel is an emerging solution to overcome these issues. NEG acts as an effective solution to deliver lipophilic drugs via topical administration. In this study, six NEs and NEGs were optimized by using a Design of Experiment (DoE) study. Optimized formulations exhibited suitable physical and chemical properties such as small and uniform particle size, skin compatible pH and desirable spreadability. NEs displayed shear thickening properties and NEGs showed shear thinning properties which are suitable for topical application. A well dispersed nanoemulsion system was confirmed by transmission electron microscopy. Drug release profile of the formulations was influenced by gel matrix, drug and excipient interaction besides molecular diffusion. The hemolysis study proved the nanoemulgels to be non-hemolytic. Antibacterial study and rabbit skin irritation test further solidified the claim of the formulations as a safe, effective and patient compliant topical treatment for various skin infections. In this research study, the developed formulations represent a promising approach for enhanced topical delivery of ciprofloxacin to combat infectious diseases and provide improved therapeutic potency.
Fabrication of In Situ Gelling Decellularized Human Amniotic Membrane Microgels for Chronic Wound Healing
Student: Fabliha Noshin
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Human amniotic membrane based novel microgels were fabricated by in situ gelling by ionic crosslinker, followed by lyophilization and ball milling. Collagen fragments along with other extracellular matrix proteins and growth factors derived from decellularized human amniotic membranes were entrapped and bound with H-bonds with sodium alginate and carboxymethyl cellulose in each microgel. Ionic crosslinking of these polymers vary with crosslinker concentration which varies the mechanical properties of the 3D scaffold system of polymers and ECM compounds. Concentration of crosslinker forming appropriate mechanical property of the 3D scaffold that can be lyophilized to a brittle structure that can be ball milled to fine gel particles in the size range of micrometers was determined. Microgels were high swelling as their surface to volume ratio is higher than that of bulk hydrogels. Thus, it can absorbs excess wound exudate containing ECM degrading matrix metalloproteinase and facilitates chronic wound healing by promoting proper cell infiltration, macrophage polarization, TGF and other growth factor regulation. Decellularized human amniotic membrane microgels showed excellent biocompatibility in vitro and proper skin attachment. Reduced number of inflammatory cells were observed in diabetic mice punch wound healing trials for these ECM based microgels compared to the control groups. Commercially available collagen based healing powder SkinColoFiber was also compared to its healing capability. Results shows faster healing and reduced scar by decellularized human amniotic membrane microgels compared to SkinColoFiber. Reduction of inflammation, scar-less wound-healing, excellent wound coverage and extrudability made these novel decellularized human amniotic membrane microgels a proper replacement of membrane based scaffolds for chronic wound healing.
Platelet Lysate Incorporated Sodium Alginate Nanogel with Mussel-Inspired Chemistry for Diabetic Wound Healing
Student: Labiba Islam Salsabil
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Diabetes affects 537 million people globally, with diabetic wounds being a major complication. Around 34% of diabetics develop foot ulcers, and 15–20% of these result in amputation. Healing is often delayed due to prolonged inflammation, impaired neovascularization, and persistent oxidative stress. Platelet lysate, derived from human blood, is rich in growth factors like EGF, VEGF, PDGF, and TGF-β1, which promote wound healing. However, its short half-life leads to rapid degradation and instability, requiring frequent administration, which raises concerns about cost, toxicity, and patient discomfort, especially with invasive delivery methods. To overcome these limitations, we developed a controlled growth factor delivery system using sodium alginate nanoparticles with a non-fouling surface, achieved through PEGylation via mussel-inspired chemistry, and further entrapped in a carbomer gel. This system offers high encapsulation efficiency (~99%), controlled and sustained release (99% over 11 days), and long-term stability, making it suitable for clinical use. We developed four formulations in a stepwise manner, conducting studies on each to evaluate their efficacy. The final formulation demonstrated excellent biocompatibility (Hemolysis ~0.643 ± 0.363%, BCI ~14 ± 0.2345%) and significantly enhanced wound healing in a diabetic mouse model compared to PL alone, with superior regeneration properties. Furthermore, the gel’s favorable pH range (4–6) and excellent spreadability improve patient compliance, positioning it as a promising non-invasive alternative to existing PL delivery methods. These findings highlight the potential of our formulation in advancing diabetic wound healing and lay the foundation for developing next-generation diabetic wound healing treatments.
Antimicrobial Peptide Stabilized Chitosan/Alginate Hydrogel Layered Hemostatic Gauge Dressing with Antibacterial Activity
Student: Ahana Jyoti Ahmed
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Antibiotic resistance poses a serious threat to global health, complicating infection management and elevating mortality rates. Additionally, uncontrolled hemorrhage appears to be a primary cause in traumatic injuries and surgical procedures. This study investigates the development and evaluation of a novel, bilayer hemostatic hydrogel dressing that can serve dual purposes of antibacterial efficacy and hemorrhage control. The use of modified cotton gauze as a substrate, layered with chitosan and alginate to form a hydrogel structure, not only maintains a moist wound environment but also offers high flexibility, making it suitable for irregular organ structures. As a result of which, the dressing should be applicable to both surface bleeding and junctional bleeding without causing potential risk and improved mechanical properties. The synergistic performance of the hydrogel’s superior fluid adsorption rate and chitosan’s exceptional adhesion properties enhances hemostatic effectiveness by rapidly absorbing blood and stabilizing clot formation, facilitating the hemostasis pathway. Furthermore, successful integration and stabilization of antimicrobial peptides being a natural source of antibiotics presents a promising alternative to conventional antibiotic drugs demonstrating significant antibacterial efficacy against broad- spectrum gram-positive and gram-negative and pathogenic and non-pathogenic bacteria without any adverse effect. The fabricated dressing undergoes comprehensive in vitro physicochemical characterization and in vivo preclinical evaluations to assess its structural integrity, therapeutic efficacy and desired functionality.
Histoscope: Breast Cancer Histopathology Image Analysis using a Multi-Stage Deep Learning Approach
Students: Samiha Jainab, Natalia Raj
Supervisor: Dr. Taufiq Hasan, Professor
Abstract:
Breast cancer is one of the most common and life-threatening diseases globally. It is very important to detect it early and correctly for effective treatment. Histopatho- logical diagnosis by microscope examination of tissue samples is the gold standard but it involves a highly complex and time-consuming procedure requiring highly ex- perienced pathologists. There’s an urgent need for automated tools that can assist in making faster and more reliable clinical decisions because of the increasing number of breast cancer cases and variations in the interpretation between experts. This study explores a multi-stage deep learning approach to improve the clas- sification of breast cancer histopathology images.By refining feature extraction at multiple levels, the framework enhances the model’s ability to distinguish between different subtypes of breast cancer tissue.This structured, multi-step method allows the model to more precisely model complex tissue architecture and creates a more valid and comprehensive examination. The primary aim is to increase the reliability of our diagnoses by eliminating error, reducing false positives and false negatives, and enhancing the credibility of AI classification systems. One of the key areas of this research is developing a dataset particularly designed for the Bangladeshi population. Laboratory procedures and staining protocols can dramatically influence histopathological images, and demographic variations can similarly make AI models developed on datasets from other populations less trustworthy. To address this, we are collaborating with the National Institute of Cancer Research & Hospital (NICRH) to get and process histopathology breast cancer images of local patients from the region. This local dataset represents the regional population and makes it more versatile and clinically relevant for artificial intelligence (AI). This research combines a stringent deep learning methodology with a database specific to a target population. This enhances AI-assisted diagnosis in histopathology by reducing the burden on pathologists and also making the diagnosis more reproducible. In the future, we plan to supplement more data, investigate new deep learning algorithms, and test the model in real clinical settings. And ultimately, our objective is to develop an effective and scalable diagnostic platform that improves detection of breast cancer, facilitates early diagnosis, and allows for personalized treatment planning—enhancing patient care and outcomes.
An End-to-End Deep Learning Framework for Bone Mineral Density Prediction from X-ray Images
Students: Sayeed Sajjad Razin, Farihin Rahman
Supervisor: Dr. Taufiq Hasan, Professor
Abstract:
Bone mineral density (BMD) estimation from X-ray images is essential for early osteoporosis detection and risk assessment. However, there is a significant gap in research focused on the Southeast Asian (SEA) demographic, with no existing studies specifically tailored to this population. Furthermore, existing methods fail to utilize all available information, such as patient metadata and other BMD report information, which are crucial for improving predictive performance. This task is further hindered by the lack of publicly available datasets and the limitations of existing classification- based approaches, which often rely solely on image-based features without integrating complementary patient-specific information. In this work, we introduce a comprehensive 3-step end-to-end deep learning framework that addresses these challenges by integrating multimodal data sources—including patient history and prior BMD reports—alongside X-ray images. Additionally, we propose a novel multidirectional adaptive contrastive loss function to enhance feature representation. One of the major challenges in BMD prediction is the absence of large-scale, publicly available datasets. Existing datasets often focus on binary osteoporosis classification, which does not provide the necessary granularity for precise BMD estimation. To address this, we curated a dataset comprising X-ray images, patient metadata, complete BMD score for all the region from DEXA scan reports. The X-ray images were annotated and segmented under the close supervision of the collaborating radiologists. The dataset was preprocessed to ensure high-quality inputs, including image normalization, augmentation, and alignment techniques. Metadata including patient tabular data, such as age, gender, weight, menopausal status and clinical history, was incorporated to enhance prediction accuracy. Our proposed model is based on an end-to-end deep learning pipeline utilizing EfficientNetV2-S as the encoder backbone. This choice was guided by extensive baseline experiments conducted on publicly available osteoporosis datasets, where EfficientNetV2-S demonstrated superior performance in feature extraction. To enhance the model’s ability to capture fine-grained features, we integrated a segmentation-guided mechanism, enabling the network to focus on critical anatomical regions. Moreover, the model effectively combines image-based features with tabular patient data, allowing for a more robust and interpretable prediction framework. A core contribution of this study is the introduction of the Multidirectional Adaptive Contrastive Learning (MDACL) loss, designed to refine the latent feature space by leveraging both imaging and non-imaging data. Unlike conventional loss functions, MDACL encourages representations that account for both structural and contextual patient information. By enforcing intra-class compactness and parameter-based separation within the latent space, MDACL enhances feature discrimination, ultimately leading to improved prediction performance. Our extensive experiments demonstrate that the proposed model significantly outperforms traditional approaches. The inclusion of segmentation masks and patient metadata leads to a substantial reduction in root mean square error (RMSE) and mean absolute error (MAE), while achieving higher R2 and Pearson correlation coefficient (PCC). These results highlight the effectiveness of the MDACL loss function in improving the generalisation ability of deep learning models for BMD prediction. In conclusion, this work presents a novel deep learning approach that successfully integrates imaging and tabular data using an advanced contrastive loss function. The findings pave the way for future research in leveraging multimodal data for precision medicine applications, with potential extensions to broader medical imaging tasks.
Development of PEG-Modified HPMC-Pluronic Composite Film for Enhanced Drug Delivery in Drug-Coated Balloon Therapy
Students: Zarin Subah, Hridoy Sen Munna
Supervisor: Dr. Jahid Ferdous, Associate Professor
Abstract:
Neointimal hyperplasia is a prevalent consequence of arterial damage following interventions such as angioplasty, stenting, or vascular grafting. The phenomenon is characterized by the proliferation and migration of vascular smooth muscle cells (VSMCs) and the accumulation of extracellular matrix in the intima. This pathological remod- eling can occur, which often results in restenosis. Restenosis is characterized by the accumulation of smooth muscle cells around the stent [1]. Drug-coated balloon (DCB) therapy has emerged as a viable technique for administering antiproliferative drugs during short balloon inflation periods. However, existing DCB technologies are limited by ,tracking phase drug loss, coating stability, and suboptimal drug transfer ([2],[3]). To overcome these limitations, we developed a novel polyethylene glycol (PEG)-modified hydroxypropylmethylcellulose (HPMC) and pluronic composite film, which is optimized to enhance drug delivery in drug-coated balloon (DCB) applications. We methodically examined two principal formulation variables: the concentration of HPMC, the compar- ison of Pluronic:HPMC ratios of 2:1 against 1:1, and the inclusion of 3% w/v PEG 400. We used Fourier transform infrared (FTIR) spectroscopy to analyze the incorporation of the components in the film-based coating and conducted disintegration testing to analyze the composite films under conditions analogous to a physiological environment. We performed ex vivo tissue uptake tests on bovine artery tissues to replicate clinical drug-coated balloon deployment and evaluate drug transfer efficacy. The PEG-modified films increased the drug transfer efficiency by 13-29% in both types of formulations. The best formulation had a lower concentration of HPMC (Pluronic:HPMC = 2:1) and 3% PEG. It was able to exhibit a drug transfer of 47.65 μg / g of tissue, which is comparable to that of commercial DCB systems. ( In contrast, formulations with higher HPMC content demonstrated a significant reduction of 45-52% in drug transfer efficiency, despite the addition of PEG. However, all of the formulations exceeded the minimal tissue concentra- tion of 12−20 micrograms/g of tissue required for clinical efficacy. The addition of PEG accelerated the disintegration rate of films, which aligns with the brief inflation durations often observed in current DCB therapy [4], [5]. PEG-modified HPMC-Pluronic composite films significantly enhance drug transfer and disintegration rate in DCB applications. This technique may address certain issues with existing DCB technologies and yield improved outcomes for neointimal hyperplasia and restenosis by enhancing the polymer composition, primarily through reducing HPMC content and incorporating PEG.
A Multi-Component Finite Element Model of the Human Eye to Assess Glaucoma Risk under Pathological Conditions
Student: Mollah Adib Aktab
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
This study introduces a multi-component finite element model to investigate the biome- chanical behavior of the optic nerve head (ONH) under pathological conditions, particularly focusing on the effects of intraocular pressure (IOP), intracranial pressure (ICP), and ocu- lar pulse amplitude (OPA). The research aims to explore the complex interactions between these factors and their impact on the deformation, strain, and stress of the lamina cribrosa (LC) and post-laminar neural tissue (PLNT), critical components in glaucoma development. The model incorporates detailed geometrical and material properties of the eye, includ- ing adipose tissue, choroid, retina, and sclera, to provide a more realistic representation of ocular structures. Through numerical simulations, the study analyzes the effect of varying IOP, ICP, and OPA on the optic nerve head, with a particular focus on understanding how these factors contribute to optic nerve damage and glaucoma progression. The results demonstrate that increased IOP significantly deforms the lamina cribrosa, with corresponding increases in shear stress and strain. Furthermore, variations in ICP and OPA also influence LC deformation, highlighting their importance in glaucoma risk. Mate- rial properties such as the Young’s modulus of the lamina cribrosa play a crucial role in the deformation response, with higher values leading to greater resistance against deformation, thus affecting the glaucoma risk assessment. This study provides a deeper understanding of the biomechanical mechanisms under- lying glaucoma, offering insights into how variations in IOP, ICP, and OPA contribute to ocular tissue deformation. It also presents a novel approach by integrating blood flow mod- eling into the finite element analysis, enhancing the accuracy of glaucoma risk assessments and providing valuable data for future clinical applications and experimental validation.
Analysis of Sodium Alginate-Based Nanoformulations for the Treatment of Ophthalmologic Infection
Student: S. M. Mursalin Sonet
Supervisor: Dr. Muhammad Tarik Arafat, Professor
Abstract:
Two sodium alginate-based nanoformulations were prepared to treat bacterial infections. In the semi-gel formulation, 1% sodium alginate was used as a polymer, and 1% CaCl2 was added as a crosslinker to 5 ml of drug solution. The drug used was 0.5% moxifloxacin (antibiotic). The pH of the solution was initially 5.0; however, since the human eye has a pH of 7.0-7.4, the pH was adjusted to this range using 2 ml of NaOH to stabilize the solution. The solution was then sonicated for 5 minutes. In the nanosuspension formulation, 0.5% sodium alginate and 0.5% CaCl2 were used as the polymer and crosslinker, respectively, in 5 ml of drug solution. To reduce viscosity and enhance stability, 1% Pluronic F-127 was added. The pH of this solution was also adjusted to the eye’s pH range using 2 ml of NaOH. To determine which formulation performed best, several in vitro and in vivo tests were conducted and compared with the commercial drug, Vigalon. In vitro tests included measuring the zeta potential and particle size; both formulations showed values greater than ±30 mV, indicating high stability, and particle sizes were at the nano-level. Mucoadhesion tests showed that the nanosuspension outperformed both the semi-gel and the commercial drug across various sample amounts. The hemolysis ratio was less than 5%, indicating that the formulations are non-hemolytic. In antibacterial tests against pathogenic S. Aureus, the nanosuspension produced the largest zone of inhibition (ZOI). In vivo tests in rabbits showed that the nanosuspension treated infections in just three days, whereas the semi-gel and Vigalon took four and five days, respectively. Intraocular pressure (IOP) began to stabilize from day two with the semi-gel and the commercial drug, whereas it took three days with the nanosuspension. Dosing tests revealed that when applied four times a day, the nanosuspension cured the infection in just four days, compared to five days with the same regimen for the commercial drug. In terms of bioavailability, however, the commercial drug Vigalon outperformed both nanoformulations.