Abstract
Automated diagnosis of Chest X-rays (CXR) is a critical goal in medical imaging. SwinCheX introduces a multi-label classification model that replaces traditional Convolutional Neural Network (CNN) backbones with a hierarchical Swin Transformer. By leveraging a 3-layer Multi-Layer Perceptron (MLP) head and adhering to a rigorous patient-wise evaluation protocol on the massive ChestX-ray14 dataset, the model achieves a new state-of-the-art (SOTA) average AUC of 0.810.
Introduction
Automated CXR diagnosis is essential for fast, reliable screening. While CNNs like ResNet and DenseNet have been standard, they often struggle with long-range dependencies and multi-scale features. Furthermore, the field faces a "fairness" crisis due to inconsistent data splits where images from the same patient appear in both training and testing sets (data leakage), leading to inflated performance metrics. SwinCheX addresses these issues with a more powerful architecture and a standardized benchmarking process using the official patient-wise split of the ChestX-ray14 dataset, which contains 112,120 images from over 30,000 patients across 14 pathologies.
Methods
The SwinCheX architecture departs from the CNN-heavy status quo by using a Swin Transformer (Swin-L) as the backbone.
Architectural Components
- Swin Transformer Backbone: Unlike the original Vision Transformer (ViT), the Swin Transformer uses a hierarchical approach with "shifted windows." This captures multi-resolution features—essential for detecting both large-scale heart issues and tiny lung nodules—while maintaining linear complexity relative to image size. The backbone is pre-trained on ImageNet-22k and fine-tuned on medical data.
- Multi-Task MLP Heads: The model employs 14 distinct MLP heads, one for each pathology. Experiments with various depths revealed that a 3-layer head (384, 48, and 48 neurons) provided the optimal balance of complexity and performance.
Results and Discussion
SwinCheX achieved a mean AUC of 0.810, surpassing the previous best of 0.799 when evaluated under the same fair, patient-wise protocol.
Key Insights
- Hierarchical Features: The Swin Transformer significantly outperformed the standard ViT (0.810 vs 0.779 AUC), suggesting that hierarchical spatial representations are vital for medical imaging where pathology sizes vary.
- Impact of the MLP Head: A "headless" Transformer using raw output tokens is insufficient. Adding a 3-layer MLP head allowed the model to learn more complex, disease-specific functions, boosting the mean AUC from 0.804 to 0.810.
- Standardized Evaluation: Re-evaluating the previous SOTA (DNet) using strict patient-wise protocols showed a performance drop from 0.807 to 0.799, highlighting the impact of data leakage in prior research.
- Explainability: Using Grad-CAM saliency maps, the authors confirmed the model focuses on correct anatomical regions, such as heart boundaries for Cardiomegaly, rather than image artifacts or shortcuts.
Conclusion
SwinCheX sets a new benchmark for CXR diagnosis by combining hierarchical transformers with robust evaluation standards. While successful, limitations include label noise in the ChestX-ray14 dataset (~90% accuracy) and image artifacts like watermarks. Future work should focus on specialized pre-processing to remove non-clinical features and optimizing computational costs for resource-constrained clinical settings.
Summary updated