Research Article
Dual-Stream Feature Synergy and Prior Fusion Method for Chest X-ray Anomaly Detection
@INPROCEEDINGS{10.4108/eai.22-5-2026.2365105, author={Chenxi Gao}, title={Dual-Stream Feature Synergy and Prior Fusion Method for Chest X-ray Anomaly Detection}, proceedings={Proceedings of the 4th International Conference on Image, Algorithms, and Artificial Intelligence, ICIAAI 2026, 22-24 May 2026, Singapore, Singapore}, publisher={EAI}, proceedings_a={ICIAAI}, year={2026}, month={8}, keywords={Anomaly detection; Zero-shot learning; Dual-stream network; DINOv3; Feature fusion}, doi={10.4108/eai.22-5-2026.2365105} }- Chenxi Gao
Year: 2026
Dual-Stream Feature Synergy and Prior Fusion Method for Chest X-ray Anomaly Detection
ICIAAI
EAI
DOI: 10.4108/eai.22-5-2026.2365105
Abstract
The current single-stream unsupervised anomaly detection algorithms focus on two problems which are: CNNs (Convolutional Neural Networks) ignoring global context and pre-trained ViT (Vision Transformer) models ignoring layered deep features and losing fine grained pathological textures. To remedy this, the study proposes a dual-stream synergy architecture framework which integrates both ‘texture’ and ‘context’ streams. More precisely, the texture branch employs ResNet50 local features at high frequencies and the context branch integrates anatomy prior filters, background noise suppression, and multi-level aggregation from DINOv3 to extract texture at shallow and middle levels. The Kermany Chest X-ray database experiments show that this supplementary feature synergy improves detection performance. The dual-stream model (AUC=0.8944) indeed outstrips DINOv3 (AUC=0.65) and ResNet50 (AUC=0.8728) standalone models, cementing the practicality of the CNNs inductive bias and the Transformers global semantics to achieve strong zero-shot medical diagnosis.
