pathmnist

Based on our paper "Implementing vision transformer for classifying 2D biomedical images" published in Scientific Reports (Nature)

🔍 Enhance medical imaging with a lightweight CNN model that offers over 91% accuracy and integrated explainability for better clinical trust.

Medical Image Synthesis project (MedSyn). In-depth evaluation of the efffects of different synthesis models (i.e., CFG ccDDPM) for medical image synthesis for class balancing on image datasets (i.e., PathMNIST).

A lightweight Explainable AI CNN for PathMNIST medical imaging, achieving 91%+ accuracy with Integrated Gradients and SQLite-based attribution storage. Built in PyTorch, this scalable model delivers high performance, transparency, and real-world readiness, making it ideal for medical AI, edge deployment, and explainable deep learning research.

This script focuses on explaining a pre-trained CNN’s predictions on the MedMNIST dataset using Deep SHAP. Shapley values are computed for each pixel, summed to create a single Shapley score per image, and saved alongside labels in CSV files for interpretability in medical image classification.

Reproducible PathMNIST reliability benchmark: imbalance, calibration, external-centre stress tests, failure slices, and explanations.

IRSANet Case Study: Patch-level classification of histopathological colon tissue images

A comparative study and implementation of five autoencoder architectures (Basic, CNN, Denoising, VAE, Sparse) for medical image reconstruction and feature extraction using the PathMNIST dataset.

PathMNIST classification with PyTorch (DirectML) and ONNX

CNN, MLP and Random Forest comparison for PathMNIST histopathology image classification.

PyTorch experiments for PathMNIST classification and MNIST knowledge distillation using compact CNNs and ResNet18.