This repository is a part of my final year project in the domain of Deep Learning, Explainable AI, and AI in Healthcare submitted to the National University of Sciences & Technology (NUST). The project details are given below:
OptiGuard.mp4
Glaucoma is a leading cause of blindness that requires early diagnosis. This project introduces a generalized, attention driven & explainable system for early glaucoma classification using retinal fundus images via Convolutional Neural Networks (CNNs).
Timely & accurate glaucoma detection is a challenge. Manual methods cause treatment 3. delays. This project aims to address this need with an automated explainable CNN system.
-
Utilizes the G1020 dataset for training. RFIs are preprocessed and augmented for variability. Detectron2 with Mask RCNN and ResNet-50 backbone is employed for OD and OC segmentation, validated with Average Precision metrics. Computes diagnostic metrics like CDR and NRR area.
-
Uses the SMDG-19 dataset, preprocessed with resizing, normalization, and augmentation including Histogram Equalization and CLAHE. Focuses on OD and OC regions identified by the segmentation module. Trains EfficientNet-B0 with Cross-Entropy loss and Adam optimizer, addressing class imbalance with weighted sampling.
-
Finally, Grad-CAM, Attention Mechanisms and LLM are employed for result explainability and interpretability.
- Python, PyTorch, Keras, Scikit-Learn, Pandas, Numpy, Matplotlib, Seaborn
- Detectron2
- Flask
- G1020 Dataset : https://www.kaggle.com/datasets/arnavjain1/glaucoma-datasets
- SMDG Dataset : https://www.kaggle.com/datasets/deathtrooper/multichannel-glaucoma-benchmark-dataset
This project was successfully completed with contributions from: - Syed Safi Ullah Shah (Safi50) - Humza Khawar - Muhammad Huzaifa and under the advisory of: Dr. Muhammad Naseer Bajwa (https://scholar.google.com.pk/citations?user=PeeIGEgAAAAJ&hl=en)