DAAP-NET: Automatic identification and quantitative analysis of gastric wall structure for cancer screening using gastric ultrasound images

Abstract

Objectives

Gastric cancer remains a significant public health challenge worldwide, ranked fifth in incidence and fourth in mortality among all malignant tumors. Early gastric cancer (EGC) detection is critical to improving survival rates.

Methods

We propose a new segmentation model, Dual Attention Atrous Pixel Network (DAAP-Net), designed to predict the gastric wall and stomach cavity in gastric ultrasound images. Post-segmentation, the predicted gastric wall mask extracts a region of interest (ROI) focused on the five-layer wall structure. The ROI undergoes a spatially diffused iterative enhancement (SDIE) technique to suppress intra-layer noise while preserving inter-layer transitions. We apply edge detection to the SDIE-refined ROI and compute layer thickness as pixel distances between successive edges, and normalize them into the proportion vector $\text{x}$. A scalar deviation $\text{d}$ from the normal baseline quantifies abnormality.

Results

DAAP-Net outperforms state-of-the-art segmentation methods, achieving Intersection over Union scores of 0.7720 ± 0.0618 for normal gastric wall, 0.9007 ± 0.0495 for normal stomach cavity, 0.7607 ± 0.0780 for cancer gastric wall, and 0.8843 ± 0.0561 for cancer stomach cavity. Quantitative analysis shows gastric wall layer parameters differ markedly; the edge-derived deviation metric $\text{d}$ separates cohorts, with normal mean 0.128 and cancer mean 0.508.

Conclusions

Our research highlights structural differences between normal and cancerous gastric walls, providing a reliable and non-invasive method for EGC detection. Current limitations include manual ROI selection and occasional errors in low-contrast regions. Future work includes automated ROI selection, adding a benign-labeled cohort, a multi-center dataset, and improving model accuracy for real-time clinical applications.