MedicalPatchNet: A Patch-Based Self-Explainable AI Architecture for Chest X-ray Classification

MedicalPatchNet is a novel, inherently self-explainable deep learning architecture for chest X-ray classification that achieves performance comparable to EfficientNetV2-S while significantly improving diagnostic interpretability and pathology localization accuracy through a transparent patch-based aggregation mechanism, thereby enhancing clinical trust and safety.

The Gist

The Big Problem: The "Black Box" Doctor

Imagine you have a brilliant new AI assistant that can look at a chest X-ray and tell you if a patient has pneumonia, a broken rib, or a heart issue. It's incredibly accurate—sometimes even better than human doctors.

But here's the catch: It won't tell you why.

It's like a magic 8-ball that gives you the right answer but refuses to explain its reasoning. If the AI says, "This patient has pneumonia," a doctor might ask, "Where? Is it in the left lung or the right? Is it near the heart?" The AI just says, "Trust me, I'm smart."

In medicine, you can't just trust a black box. If the AI is wrong, or if it's "cheating" by looking at the wrong thing (like a label on the X-ray saying "Left" instead of the actual lung), the doctor needs to know so they don't make a mistake.

The Solution: The "Mosaic" Approach (MedicalPatchNet)

The researchers created a new AI called MedicalPatchNet. Instead of looking at the whole X-ray as one giant picture, they changed how the AI "sees" the image.

The Analogy: The Jigsaw Puzzle
Imagine the X-ray is a giant jigsaw puzzle.

• Old AI (The "Black Box"): Looks at the whole puzzle at once. It guesses the answer based on the whole picture, but you can't see which specific pieces mattered.
• MedicalPatchNet: Takes the puzzle apart. It looks at one small square piece at a time.
  1. It looks at Piece #1 and asks: "Does this piece look like pneumonia?"
  2. It looks at Piece #2 and asks: "Does this piece look like pneumonia?"
  3. It does this for every single piece in the puzzle.
  4. Finally, it takes all those little answers and averages them to get the final diagnosis.

Why is this better?
Because the AI made its decision piece-by-piece, we can see exactly which pieces voted "Yes" and which voted "No."

• If the AI says "Pneumonia," we can look at the "heat map" and see that the red (positive) votes came from the bottom of the lung.
• If the AI is cheating and looking at a text label on the side of the image, the "red votes" will show up on the text, not the lung. The doctor can immediately spot the cheat and ignore the AI's advice.

How It Works (The "Voting Booth")

Think of the X-ray as a town with 64 neighborhoods (patches).

1. Independent Voters: Each neighborhood has its own little expert who votes on whether the patient is sick. They don't talk to each other while voting; they only look at their own street.
2. The Count: At the end, the mayor (the AI) counts all the votes.
3. The Map: Because the votes are counted separately, the mayor can draw a map showing exactly which neighborhoods voted "Sick" and which voted "Healthy."

This is called "Self-Explainable." The AI doesn't need a special tool to explain itself later; the explanation is built right into how it thinks.

Did It Work?

The researchers tested this new AI on a massive database of over 220,000 chest X-rays.

• Accuracy: It was just as good as the best existing AI (called EfficientNet). It didn't sacrifice accuracy for transparency.
• Honesty: When they tested if the AI could point to the actual disease (localization), MedicalPatchNet was much better than the old methods.
  • Old Method (Grad-CAM): Like a spotlight that sometimes shines on the wrong spot or is too blurry.
  • MedicalPatchNet: Like a laser pointer that hits the exact spot where the disease is.

The "Shortcut" Test

One of the coolest parts of the paper is how it catches AI "cheating."
Sometimes, AI learns shortcuts. For example, if all the "Pneumonia" X-rays in the training data had a little "R" (for Right) written on them, the AI might learn that "R = Pneumonia" instead of looking at the lungs.

Because MedicalPatchNet looks at pieces individually:

• If the AI relies on the "R," the red votes will appear on the letter "R."
• If the AI relies on the lung, the red votes appear on the lung.
• The doctor can see this immediately and know, "Oh, this AI is looking at the wrong thing."

The Bottom Line

MedicalPatchNet is a new way of building medical AI that is honest by design.

Instead of being a mysterious genius that gives answers without reasons, it acts like a team of specialists, each looking at a small part of the picture and reporting their findings. This allows doctors to trust the AI, understand its reasoning, and catch it if it tries to take a shortcut. It's a big step toward making AI a safe, reliable partner in the hospital.

Technical Summary

1. Problem Statement

Deep neural networks (DNNs) have achieved state-of-the-art performance in radiological image classification but suffer from a lack of interpretability, often operating as "black boxes." This limits their clinical adoption because:

• Post-hoc limitations: Existing explainability methods (e.g., Grad-CAM, Eigen-CAM) are applied after training. Research indicates these methods can be misleading, highlighting regions that do not actually drive the decision, or failing to reveal "shortcut learning" (e.g., relying on image borders or laterality markers rather than pathology).
• Complexity: Post-hoc explanations often require deep technical knowledge to interpret correctly, making them inaccessible to clinical practitioners.
• Trust: Without inherent transparency, clinicians cannot verify if a model's high accuracy is based on genuine pathological features or dataset biases.

The authors argue for Self-Explainable AI (XAI): architectures designed to be interpretable by construction, rather than relying on post-hoc approximations.

2. Methodology: MedicalPatchNet

MedicalPatchNet is a novel architecture designed to be a "drop-in replacement" for standard image-level classifiers (like EfficientNet) while providing inherent, transparent explanations.

Core Architecture

• Patch-Based Decomposition: The input chest X-ray (resized to $512 \times 512$) is split into a grid of non-overlapping patches (e.g., $8 \times 8$ grid of $64 \times 64$ patches).
• Independent Processing: Each patch is processed independently by a shared backbone network (EfficientNetV2-S in this study). There is no communication between patches during the feature extraction phase.
• Fixed Aggregation: The global prediction is computed by taking the arithmetic mean of the logits (raw pre-sigmoid scores) from all patches.
  • Formula: $Z = \frac{1}{P^2} \sum_{i=1}^{P^2} z_i$, where $z_i$ is the logit vector for patch $i$.
  • The final classification is $\hat{y} = \sigma(Z)$.
• No Learnable Attention: Unlike Multiple Instance Learning (MIL) frameworks that use learnable attention mechanisms to weigh patches, MedicalPatchNet uses a fixed, non-learnable average. This ensures that the contribution of each patch to the final decision is explicit and not learned/hidden.

Explainability Mechanism

• Direct Attribution: Because the global logit is a simple sum of local logits, the contribution of any specific patch to the final decision is directly quantifiable.
• Saliency Maps:
  • Raw Patch Saliency: Visualizes the raw logits of each patch. Positive values indicate evidence for a class; negative values indicate evidence against it.
  • Scaled Patch Saliency: Raw logits are multiplied by the final classification probability ($\hat{y} \odot z_i$) to incorporate global context while maintaining local attribution.
  • Smoothing: To reduce the "blocky" appearance of patch-based maps, the authors generate multiple spatially shifted saliency maps and average them, creating a smoother heatmap without losing the underlying patch-based logic.

Training

• Dataset: Trained on the CheXpert dataset (223,414 images) using only image-level labels (no segmentation masks).
• Backbone: EfficientNetV2-S (modified for single-channel grayscale input).
• Loss: Standard binary cross-entropy on the global output.
• Constraints: The model is trained only on classification labels; it does not use segmentation data to learn where to look, ensuring the localization capability emerges purely from the architecture's constraints.

3. Key Contributions

1. Inherently Self-Explainable Architecture: MedicalPatchNet eliminates the need for post-hoc explanation methods by making the decision process transparent by design. Every patch's contribution is explicitly defined.
2. Performance-Interpretability Trade-off: The authors demonstrate that enforcing patch-wise independence and fixed aggregation does not significantly degrade classification performance compared to a standard image-level EfficientNetV2-S.
3. Superior Localization: The model outperforms standard post-hoc methods (Grad-CAM, Grad-CAM++, Eigen-CAM) in localizing pathologies, particularly in identifying the specific regions driving the decision.
4. Shortcut Detection: The architecture naturally exposes "shortcut learning." If a model relies on a border artifact or a laterality marker, the patch-based saliency map will clearly highlight that specific patch, whereas post-hoc methods might diffuse this signal or misattribute it.
5. Accessibility: The visualizations are intuitive (heatmaps of patches) and do not require deep learning expertise to interpret, making them suitable for clinical workflows.

4. Results

The study was evaluated on the CheXpert dataset (classification) and the CheXlocalize dataset (localization/segmentation).

Classification Performance

• AUROC: MedicalPatchNet achieved an AUROC of 0.907 on CheXlocalize, nearly identical to the standard EfficientNetV2-S (0.908).
• Metrics: Sensitivity, specificity, and accuracy were comparable across all 14 classes, proving that local patch processing is sufficient for high-accuracy chest X-ray diagnosis.

Localization Performance (CheXlocalize)

• Hit Rate: Defined as the fraction of cases where the pixel with the maximum saliency lies inside the ground-truth pathology mask.
  • MedicalPatchNet (Scaled): 0.485 (Mean)
  • MedicalPatchNet (Raw): 0.471
  • Grad-CAM: 0.376
  • Grad-CAM++: 0.325
  • Eigen-CAM: 0.303
  • MedicalPatchNet outperformed all post-hoc baselines in 9 out of 10 analyzed pathologies.
• Mean Intersection over Union (mIoU): When evaluating the full spatial overlap (including false positives and false negatives), MedicalPatchNet achieved the highest overall mIoU (0.069 vs. 0.052 for Grad-CAM). This indicates it is better at avoiding false positive highlights in irrelevant regions.

Limitations Observed

• Global Context: The model struggles with conditions requiring integration of distant image regions (e.g., correlating an enlarged heart with bilateral infiltrates), as patches are processed independently until the final averaging step.
• Patch Size Sensitivity: Smaller patches ($64 \times 64$) worked well for most classes but slightly underperformed on "Pneumonia" compared to larger patches, suggesting some pathologies benefit from broader context.

5. Significance and Conclusion

MedicalPatchNet represents a paradigm shift from "explainable by post-hoc analysis" to "explainable by design."

• Clinical Trust: By providing explicit, reliable visualizations of why a decision was made, it mitigates the risk of clinicians trusting models that rely on spurious correlations (shortcuts).
• Robustness: The architecture is robust across different backbones (tested with ConvNeXt) and maintains high diagnostic accuracy.
• Future Impact: The authors suggest this approach can be extended to 3D imaging (CT/MRI) and multimodal tasks, offering a scalable path toward safer, transparent AI in medicine.

The code and model weights are publicly available, promoting reproducibility and further research into inherently interpretable medical AI.