Risk Classification of Low-Resolution Whole-Slide Thumbnail Images by Multi-dimensional Feature Reconstruction with Multi-task Deep Learning Network Helps Prioritize Pathology Case Registration.

Contemporary surgical pathology workflows often prioritize slide examination based on case registry order rather than patient risk level. As a result, high-risk cases, especially those involving malignant lesions, may be unintentionally delayed, potentially affecting patient outcomes. In this study, we present an artificial intelligence (AI)-based framework designed to efficiently screen and prioritize malignant cases by analyzing hematoxylin and eosin (H&E)-stained, low-resolution thumbnail whole-slide images (TWSIs). The proposed approach includes three key components. First, image preprocessing is performed to reduce artifacts and identify the initial tissue region. Next, a multi-task deep learning network conducts both tissue segmentation and benign-versus-malignant classification. Finally, multi-dimensional feature reconstruction is utilized to improve classification accuracy. We evaluated the performance of our framework on 334 TWSI images (746 × 1632 pixels), comprising 100 benign and 234 malignant cases. The system achieved an average inference time of 2.33 ± 0.31 s per image, along with an accuracy of 91.91%, a sensitivity of 93.59%, a specificity of 88.00%, a positive predictive value of 94.84%, and a negative predictive value of 85.56%. These results correspond to a 6.41% false negative rate. The findings suggest that applying AI-driven analysis to TWSIs can effectively expedite case triage, thereby enhancing the sorting and prioritization of surgical pathology specimens and potentially improving clinical decision-making.