CXNet-m1: Anomaly Detection on Chest X-Rays With Image-Based Deep Learning
Problem
Automatically detecting anomalies in chest X-ray images via deep learning suffered from over-fitting and low transfer efficiency when fine-tuning existing networks.
Why it matters
Faster, more accurate chest X-ray reading could improve diagnosis and work efficiency.
What the AI did
A hierarchical CNN (CXNet-m1), shorter and thinner than fine-tuned networks, was designed with a novel sin-loss function that learns discriminative information from misclassified and indistinguishable images, plus optimized convolutional kernels.
Results
Better accuracy rate, recall rate, F1-score, and AUC than fine-tuning baselines; absolute values not stated in source
Impact
Shows designing a proper CNN can outperform fine-tuning deep networks on ChestX-ray14.
Record
| Venue | IEEE Access |
|---|---|
| Published | 2018 |
| Work type | article |
| Authors | Shuaijing Xu, Hao Wu, Rongfang Bie |
| Citations | 122 (OpenAlex, at import time) |
| References | 65 |
| Model or method | CXNet-m1 (hierarchical CNN with sin-loss) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1109/access.2018.2885997 |
| Open access | gold |
| Keywords | chest X-ray, CNN, anomaly detection, deep learning, sin-loss, ChestX-ray14 |
Importance rating
real but narrower advance
Abstract calls automated chest X-ray anomaly detection an urgent need for diagnosis accuracy and work efficiency, but gives no scale; advance is a better CNN design validated on one benchmark (ChestX-ray14).