Faithful, Interpretable Chest X-ray Diagnosis with Anti-Aliased B-cos Networks

Marcel Kleinmann, Shashank Agnihotri, Margret KeuperarXiv 2025 · DOI · arXiv

Introduction

Deep learning models have demonstrated remarkable accuracy in medical imaging, yet their "black-box" nature remains a significant barrier to clinical adoption. In high-stakes environments like radiology, understanding the reasoning behind a diagnosis is as critical as the diagnosis itself. Traditional post-hoc explanation methods, such as GradCAM, often produce coarse, blurry heatmaps that may not faithfully reflect the model's internal logic.

Inherently interpretable models like B-cos networks offer a promising alternative by aligning internal weights with input features to produce high-resolution contribution maps. However, standard B-cos architectures are plagued by "aliasing artifacts"—a grid-like "screen door" effect caused by strided downsampling. These artifacts obscure clinical details and can reduce trust in the AI's reasoning. This paper introduces anti-aliased B-cos networks designed to provide faithful, artifact-free explanations for chest X-ray (CXR) interpretation.

Comparison of explanation methods. Standard B-cos suffers from grid artifacts, while the proposed anti-aliased versions (B-cos_FLC and B-cos_BP) provide clear, faithful, and interpretable maps.

Methodology: Anti-Aliasing and Multi-Label Support

The core innovation lies in replacing standard strided convolutions with a principled downsampling approach that adheres to the Nyquist-Shannon sampling theorem.

Principled Downsampling

To eliminate aliasing, the authors decouple the convolution and downsampling steps. Instead of a single strided convolution, the model applies a stride-1 convolution followed by an anti-aliasing pooling layer. This ensures that high-frequency noise does not corrupt the explanation maps. The researchers evaluated two primary methods:

  1. BlurPool: Applies a spatial blurring filter before downsampling.
  2. FLCPooling: Utilizes a low-pass filter in the frequency domain to remove high-frequency noise.
Architectural modification: replacing standard strided convolutions with a sequence of stride-1 convolution, MaxPool, and an anti-aliasing pooling layer (BlurPool or FLCPooling).

Multi-Label Extension

Medical diagnoses are rarely mutually exclusive. The authors extended the B-cos framework to handle multi-label classification, allowing the network to generate independent, high-resolution contribution maps for multiple co-occurring conditions (e.g., pneumonia and pleural thickening) simultaneously. This makes the model significantly more viable for complex clinical cases.

Key Results and Insights

The models were evaluated on the RSNA Pneumonia Detection Challenge and the VinBigData Chest X-ray Abnormalities Detection dataset. The latter features 14 thoracic conditions annotated by 17 board-certified radiologists, providing a robust ground truth for localization.

Performance and Faithfulness

The anti-aliased B-cos models achieved diagnostic performance (AUC ~88%) comparable to standard ResNet50 "black-box" models, proving that interpretability does not require sacrificing accuracy. Quantitatively, the proposed B-cos_FLC model significantly outperformed existing methods on the Energy-based Pointing Game (EPG) metric, which measures how much "explanation energy" falls within the correct bounding box. B-cos_FLC achieved a score of 28.25%, more than doubling the performance of standard B-cos (~11%) and surpassing post-hoc methods like LayerCAM (~23%).

Qualitative Clarity

Visual comparisons show a dramatic difference. Where standard B-cos maps are noisy and grid-like, the anti-aliased versions provide sharp highlights that clearly delineate pathologies. In peak recall configurations, the B-cos_FLC model reached an EPG of 91.56%, indicating that nearly all of the model's positive evidence was correctly located within the diseased regions.

Conclusion

By integrating advanced pooling techniques and multi-label support, anti-aliased B-cos networks bridge the gap between high-performance deep learning and clinical interpretability. These models provide radiologists with sharp, faithful, and artifact-free visual evidence, fostering the trust necessary for AI integration into medical workflows. Future work will explore scaling these techniques to transformer-based architectures and further refining localization through spatial priors.

Summary updated

Abstract

Faithfulness and interpretability are essential for deploying deep neural networks (DNNs) in safety-critical domains such as medical imaging. B-cos networks offer a promising solution by replacing standard linear layers with a weight-input alignment mechanism, producing inherently interpretable, class-specific explanations without post-hoc methods. While maintaining diagnostic performance competitive with state-of-the-art DNNs, standard B-cos models suffer from severe aliasing artifacts in their explanation maps, making them unsuitable for clinical use where clarity is essential. In this work, we address these limitations by introducing anti-aliasing strategies using FLCPooling (FLC) and BlurPool (BP) to significantly improve explanation quality. Our experiments on chest X-ray datasets demonstrate that the modified $\text{B-cos}_\text{FLC}$ and $\text{B-cos}_\text{BP}$ preserve strong predictive performance while providing faithful and artifact-free explanations suitable for clinical application in multi-class and multi-label settings. Code available at: GitHub repository (url: https://github.com/mkleinma/B-cos-medical-paper).
BibTeX
@misc{kleinmann2025faithful,
  author = {Marcel Kleinmann and Shashank Agnihotri and Margret Keuper},
  title = {{Faithful, Interpretable Chest X-ray Diagnosis with Anti-Aliased B-cos Networks}},
  year = {2025},
  eprint = {2507.16761},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.16761},
}
Figure