README.md

---
language: en
library_name: pytorch
license: mit
pipeline_tag: image-classification
tags:
- medical-imaging
- chest-x-ray
- explainable-ai
- efficientnet
- MedicalPatchNet
---

# MedicalPatchNet: Model Weights

This repository hosts the pre-trained model weights for **MedicalPatchNet** and the baseline **EfficientNetV2-S** model, as described in the paper:

**[MedicalPatchNet: A Patch-Based Self-Explainable AI Architecture for Chest X-ray Classification](https://www.nature.com/articles/s41598-026-40358-0)** (Nature Scientific Reports, 2026).  
Preprint available on [arXiv:2509.07477](https://arxiv.org/abs/2509.07477).

For the complete source code, documentation, and instructions on how to train and evaluate the models, please visit our main GitHub repository:

**[https://github.com/TruhnLab/MedicalPatchNet](https://github.com/TruhnLab/MedicalPatchNet)**

---

## Overview

MedicalPatchNet is a self-explainable deep learning architecture designed for chest X-ray classification that provides transparent and interpretable predictions without relying on post-hoc explanation methods. Unlike traditional black-box models that require external tools like Grad-CAM for interpretability, MedicalPatchNet integrates explainability directly into its architectural design.

The architecture divides images into non-overlapping patches, independently classifies each patch using an EfficientNetV2-S backbone, and aggregates predictions through averaging. This enables intuitive visualization of each patch's diagnostic contribution.

### Key Features

- **Self-explainable by design**: No need for external interpretation methods like Grad-CAM.
- **Competitive performance**: Matches the classification performance (AUROC 0.907 vs. 0.908) of EfficientNetV2-S.
- **Superior localization**: Significantly outperforms Grad-CAM variants in pathology localization tasks (mean hit-rate 0.485 vs. 0.376) on the CheXlocalize dataset.
- **Faithful explanations**: Saliency maps directly reflect the model's true reasoning, mitigating risks associated with shortcut learning.

---

## How to Use These Weights

The weights provided here are intended to be used with the code from our [GitHub repository](https://github.com/TruhnLab/MedicalPatchNet). The repository includes scripts for data preprocessing, training, and evaluation.

## Models Included

- **MedicalPatchNet**: The main patch-based, self-explainable model.
- **EfficientNetV2-S**: The baseline model used for comparison with post-hoc methods (Grad-CAM, Grad-CAM++, and Eigen-CAM).

---

## Citation

If you use MedicalPatchNet or these model weights in your research, please cite our work:

```bibtex
@article{wienholt2026medicalpatchnet,
  title={MedicalPatchNet: a patch-based self-explainable AI architecture for chest X-ray classification},
  author={Wienholt, Patrick and Kuhl, Christiane and Kather, Jakob Nikolas and Nebelung, Sven and Truhn, Daniel},
  journal={Scientific Reports},
  year={2026},
  publisher={Nature Publishing Group UK London}
}
```