Question

What is the best CNN architecture for chest X-ray classification?

March 3, 2026

University of Lahore

Publication ID: pub0000000000000000000000000000000000000003

Abstract

I am building a chest X-ray abnormality classification system for a hospital project. Currently comparing: - ResNet-50 (pretrained on ImageNet) - DenseNet-121 (pretrained on CheXNet) - EfficientNet-B4 - Vision Transformer (ViT-B/16) Dataset: 108,000 X-rays, 14 pathology classes, class imbalance is significant. Which architecture would you recommend? Any experience with weighted loss functions for imbalanced medical data?

Publication Stats

Total Reads

Recommendations

Downloads

Citations data will be displayed here

References will be listed here

Discussion (2)

Ali Khan Mar 3, 2026 · University of Lahore

Based on my experience with chest X-rays: DenseNet-121 pretrained on CheXNet outperforms ResNet on most pathologies. The dense connections help with feature reuse which is critical for medical imaging.

Sara Ahmed Mar 3, 2026 · MIT

For class imbalance, focal loss (γ=2, α=0.25) combined with oversampling of minority classes worked best in our experiments. Also try label smoothing.

Add a comment

Publication metrics

Recommendations 1

Downloads 0

Citations 0

Author

Ali Khan

University of Lahore

View Profile