Asfina Hassan Juicy and Raiyun Kabir, under the supervision of Dr. Taufiq Hasan
Project Overview
The ColpoSense project, initiated as my undergraduate thesis in July 2023, focuses on developing a deep learning-based framework for early prediction of cervical cancer and pre-cancer stages. This project addresses the pressing need for accessible cancer screening in Bangladesh, where the patient-to-doctor ratio is approximately 4000:1, with limited access to gynecological specialists. Our goal is to provide a reliable clinical decision support system that assists non-specialist healthcare providers in making reliable diagnostic decisions in underserved areas.
Research Significance
Cervical cancer screening remains a critical challenge, especially in resource-constrained settings. Our work is rooted in the recognition that access to specialized care is limited, and as a result, many women are diagnosed at late stages. By automating the complex process of colposcopic image analysis using artificial intelligence, we aim to make early screening more accessible, reducing mortality rates associated with delayed diagnosis. The project aims to empower healthcare professionals in remote areas with decision-support tools that bridge the gap between expertise and accessibility.
Key Achievements
We developed the ColpoSense framework, which includes models for classifying transformation zones and predicting Swede scores from colposcopic images, achieving an accuracy of 81.08% and compatible precision-recall metrics. The approaches address the class imbalance and integrate attention mechanisms to improve spatial feature extraction.
For Swede Score prediction, the model achieved strong predictive performance across several key metrics (MAE, MSE, RMSE) for the Aceto-White Uptake, Margins, Surface, Vessels, Lesions, and Iodine Staining scores. This enhances the model's utility in real-world diagnostic scenarios.
A major aspect of this work was the collection and curation of a small but comprehensive dataset from a local hospital, ensuring ethical protocols were strictly followed. This dataset allowed for the model’s development and continues to grow as we work toward curating a detailed, large-scale dataset. The study record has been recently approved and published by ClinicalTrials.gov under the NIH National Library of Medicine. Use the ID NCT06644248 or follow this link to view study details.
Technical Contributions
Several advanced techniques were employed to enhance model robustness, including data augmentation, SMOTE oversampling, and automated detection of faulty images, resulting in improved model generalizability.
Interdisciplinary collaboration with doctors ensured that our technical solutions aligned with real-world clinical needs. This collaboration helped tailor the AI framework to fit into existing healthcare workflows, making the solution more practical and feasible for widespread adoption.
Future Directions
The next phase involves expanding the ColpoSense dataset to create a robust AI solution that is fully developed within the country. Future work will include collaboration with medical practitioners to refine model performance and integration into clinical workflows, exploration of advanced AI models for ROI detection, and development of a clinical decision support system accessible from local hospitals.
Personal Growth and Reflections
The ColpoSense project has been a transformative learning experience, giving me firsthand insights into the challenges and potential of applying AI in healthcare. I gained valuable technical skills, particularly in deep learning, Python, and dataset curation, while learning the importance of interdisciplinary teamwork as well. This project has fueled my ambition to continue developing AI-driven solutions for healthcare, focusing on improving accessibility and diagnostic accuracy in resource-limited settings.
Presentation
We presented our work at the 19th SAARC Federation of Oncologists (SFO) & Bangladesh International Cancer Congress 2024 on December 14th, 2024.
ColpoSense_SFO_CancerCongress_final_TH.pdfShams Nafisa Ali, Asfina Hassan Juicy, and Raiyun Kabir
In collaboration with a faculty-led team, I am contributing to developing a novel deep-learning model for breast cancer detection through ultrasound image classification. So far, we have achieved 92% accuracy in classifying the normal breast (Class: 0), benign (Class: 1), and malignant (Class: 2) tumor conditions. My role involves enhancing the data pipeline, exploring parameter-tuning strategies from current journal articles, conducting model testing, and performing cross-validation. We compare our model with state-of-the-art methods and aim to prepare a manuscript after thorough testing. This project also expanded my ability in PyTorch.
Nawsabah Noor, MBBS, FCPS; Raiyun Kabir; Dr. Mahbub Mayukh Rishad, MBBS, FCPS; Mahian Kabir Joarder; Dr. Md. Mohiuddin Sharif, MBBS, FCPS; Dr. Khairul Islam, MBBS, FCPS; Prof. Robed Amin, MBBS, FCPS; Prof. Quazi Tarikul Islam, MBBS, FCPS; and Taufiq Hasan, PhD.
In February 2025, I joined the team to conduct the second phase of the clinical feasibility study of the DengueDrops app developed by our mHealth Lab at BUET. We provided a brief training to the participating physicians before calculating IV fluid requirements with and without the app. The app usage and training improved the IV fluid calculation efficiency, making it faster and more accurate. I completed the data analysis of this study for both phases.
Copy of DengueDrops at PMCH (Feb'25).pdfKabir, R., Bappa, M. N., Islam, R., Hamid, T., Ali, S. N., & Ferdous, J. (2024). AnkleGlide: A continuous ankle flexion device for long-term bedridden patients. Journal of Orthopaedic Reports, 100514. https://doi.org/10.1016/j.jorep.2024.100514
Overview
The AnkleGlide project was born out of the pressing need to prevent complications such as deep vein thrombosis (DVT) and muscle atrophy in long-term bedridden patients. This compact, affordable device provides continuous ankle flexion, facilitating blood circulation in the lower limbs. It offers an alternative to Sequential Compression Devices (SCDs), which are costly and less accessible in low-resource settings.
Features
Simulates plantarflexion and dorsiflexion of the foot using servo motors, designed to mimic walking movement and enhance calf muscle activity.
The device features an app with Bluetooth connectivity, allowing caregivers to adjust operation speed and cycle count remotely. This user-friendly interface simplifies device control and makes the device more adaptable.
Lightweight at 2.8 kg, the device measures 25.1 cm x 16.1 cm x 29.8 cm, making it portable and easy to fit into ICU settings without causing discomfort to patients.
With a production cost of around 70 USD, the device is an economical solution aimed at low- and middle-income countries, with potential cost reductions through mass production.
Impact
AnkleGlide has been validated to improve blood circulation to the lower limbs, reducing the risks of DVT and muscle atrophy. The continuous ankle motion is proven to pump venous blood back toward the heart, improving patient outcomes during prolonged immobility. Additionally, the prototype development was guided by continuous feedback from clinicians to ensure practical application in clinical settings. AnkleGlide stands as a viable option for ICU patients who are immobilized for extended periods. Its ability to stimulate calf muscles through controlled ankle movement will not only reduce the hospitalization period but also aid in the rehabilitation process.
The current iteration of AnkleGlide works effectively, but future improvements are possible. These include increasing the movement range through higher-torque servo motors, incorporating a DVT sensor for early clot detection, and adding a muscle stimulator for severe immobility cases.
The paper “AnkleGlide: A Continuous Ankle Flexion Device for Long-Term Bedridden Patients” has been accepted for publication in the Journal of Orthopaedic Reports (JOREP), Elsevier. It represents the culmination of extensive collaboration with clinicians, iterative design, and practical problem-solving aimed at addressing a critical medical need.
AnkleGlideJOREP.mp4
Recognition
Our project AnkleGlide reached the final round of the National STEAM Olympiad 2023-24, held on July 6th, 2024.
