A depth-wise separable residual neural network for PCDH8 status prediction in thyroid cancer pathological images

Accurate prediction of protocadherin 8 (PCDH8) gene expression status from whole-slide pathological images (WSIs) is critical for thyroid cancer diagnosis and prognosis, as PCDH8 overexpression is associated with tumor aggressiveness and poor outcomes. Existing methods for PCDH8 detection are often costly, time-consuming, or require specialized expertise. To address these limitations, we developed a novel depth-wise separable residual neural network (DSRNet) for noninvasive PCDH8 status prediction directly from WSIs. We collected 403 thyroid cancer WSIs from The Cancer Genome Atlas (TCGA), with PCDH8 expression status classified as high or low based on median expression values. Each WSI was divided into 512 × 512 pixel tiles, with the top 100 non-white tiles selected per slide. DSRNet integrates depth-wise separable convolutions, residual connections, and a deformable convolutional pyramid pooling (DCPP) module to efficiently capture multiscale and long-range features in gigapixel WSIs. The model was trained using tenfold cross-validation. DSRNet achieved state-of-the-art performance with 92.76% accuracy, 91.92% precision, 92.69% recall, and 0.93 AUC on the TCGA-THCA dataset, significantly outperforming leading CNN and Transformer models. Ablation studies confirmed the contributions of each component, and attention visualization showed that DSRNet focuses on biologically relevant regions. The model also generalized well to a breast cancer dataset (TCGA-BRCA), achieving 89.13% accuracy. We developed DSRNet, a deep learning-based model for predicting PCDH8 status directly from routine H&E-stained pathological images. DSRNet combines the efficiency of convolutional operations with enhanced long-range dependency modeling, providing a noninvasive, accurate, and interpretable tool for auxiliary thyroid cancer diagnosis and prognosis. The results demonstrate its strong potential for clinical translation, though further multicenter validation is warranted.