LiteXrayNet: Bilateral Asymmetry-Aware Attention for Lightweight Pediatric Pneumonia Detection

TMLR Paper7451 Authors

Abstract: Pediatric pneumonia remains a major cause of mortality among children under five, with the greatest burden in resource-constrained settings where access to timely diagnosis is limited. Although deep learning methods have achieved strong performance in chest X-ray analysis, many existing approaches rely on large models that are difficult to deploy in such environments and do not explicitly account for the bilateral anatomical structure that radiologists routinely use during interpretation. We present LiteXrayNet, a lightweight convolutional neural network that incorporates Bilateral Asymmetry Attention (BAA), a geometry-guided attention mechanism designed to model left-right lung correspondence through spatial splitting, horizontal flipping, and adaptive feature gating. With only 127K parameters, LiteXrayNet achieves competitive pneumonia classification performance, attaining an F1 score of 97.31% and an accuracy of 97.90%, while supporting real-time inference on edge hardware with latencies of 4.11 ms on GPU and 14.53 ms on CPU. Feature-level bilateral asymmetry analysis indicates that BAA induces representations that differ systematically from those produced by generic attention mechanisms, while Grad-CAM visualizations suggest anatomically structured attention patterns consistent with common radiological reasoning. These results suggest that incorporating domain-specific anatomical priors as architectural constraints can support efficient and interpretable models suitable for deployment in resource-limited clinical settings.

Submission Type: Long submission (more than 12 pages of main content)

Previous TMLR Submission Url: https://openreview.net/forum?id=yfJCllstyT&nesting=2&sort=date-desc

Changes Since Last Submission:

## Changes Since Last Submission

We are grateful for the opportunity to resubmit our manuscript. This revision presents a substantially evolved contribution with a fundamentally redesigned architecture and enhanced experimental methodology.

### Core Architectural Redesign

We have transitioned from our previous approach to introduce **Bilateral Asymmetry Attention (BAA)**, an entirely new geometry-guided attention mechanism. BAA explicitly models left-right lung correspondence through anatomical priors, motivated by how radiologists systematically compare bilateral lung fields during diagnostic interpretation.

The mechanism employs geometric operations including spatial splitting, horizontal flipping, and asymmetry computation to encode domain knowledge directly into the architecture. This represents a more clinically grounded and theoretically principled approach to attention in medical imaging.

### Novel Interpretability Framework

We introduce the **Bilateral Asymmetry Score (BAS)**, a quantitative metric for analyzing whether learned representations respect anatomical structure. This framework enables rigorous comparison between geometry-guided and generic attention mechanisms, revealing fundamental differences in how they encode bilateral relationships.

### Strengthened Experimental Validation

The empirical evaluation has been substantially enhanced:

- **Statistical rigor**: Comprehensive significance testing, confidence intervals, and corrections for multiple comparisons
- **Pareto optimality analysis**: Systematic demonstration of favorable performance-efficiency trade-offs
- **Multi-dataset generalization**: Zero-shot transfer experiments on external datasets with different patient populations
- **Calibration analysis**: Evaluation of probability calibration quality using Brier score and Expected Calibration Error
- **Unified protocol**: All baselines trained under identical conditions with individual hyperparameter optimization
- **Comprehensive ablations**: Systematic component analysis and controlled comparisons against established attention mechanisms

### Theoretical Contribution

The work articulates a generalizable principle: **encoding domain-specific anatomical priors through explicit geometric operations can simultaneously improve performance, efficiency, and interpretability**. This extends beyond pneumonia detection to any medical imaging task involving symmetric anatomy.

### Summary

This resubmission presents a new architecture grounded in clinical reasoning, a novel interpretability framework, comprehensive statistical validation, and systematic Pareto optimality analysis. We believe these improvements strengthen the contribution to efficient, interpretable medical image analysis and look forward to the reviewers' feedback.

Assigned Action Editor: ~Stanislaw_Kamil_Jastrzebski1

Submission Number: 7451