Explainable AI for histological cancer detection
October 19, 2025
Deep learning can now classify microscopic tissue images with high accuracy, but pathologists rarely trust a model that cannot show why it reached a decision. This project builds cancer-detection models that are both accurate and interpretable. It is funded by the Ministry of Science and Technology (MOST) and runs in collaboration with Bangladesh Medical University (BMU).
At a glance
- Project: Enhancing Histology Cancer Cell Detection: Deep Learning Model Interpreted with eXplainable AI (XAI) for Microscopic Tissue Image Analysis
- Status: Ongoing (started 19 October 2025)
- Funding: Ministry of Science and Technology (MOST), Bangladesh
- Collaborator: Bangladesh Medical University (BMU)
- Principal investigator: Dr. Md Mehedi Hasan
The approach
The project builds on a 2024 student research project of the same name by Shamaeta Paul Gupta and Apurba Roy. The resulting framework (preprint):
- combines an ensemble of three pretrained CNNs: ResNet50, VGG16, and InceptionV3;
- uses a 4× super-resolution step to sharpen tissue images before classification, which improved accuracy by 1–10% (ResNet50 96.21%, VGG16 92.14%, InceptionV3 96.42%);
- explains each prediction with Grad-CAM, SHAP, and LIME, so clinicians can see which regions and features drove the decision.
The MOST-funded project continues this work with clinical collaborators at BMU.
Related medical-AI work in the lab
- Explainable vision-language models for medical reports. An ongoing MPhil project (Md Julhas Hossain) is building a framework that automatically interprets diagnostic medical reports and explains them to patients.
- Bone health from X-rays. IG-CNN (IEEE ICECTE 2026) and a follow-up preprint apply information-geometric deep learning to knee osteoporosis diagnosis.
- Biosignals. FA-EEGNet classifies stroke-rehabilitation movements from EEG, and a 2025 student project explored an EEG-actuated bionic hand.