Deep learning-based prognostic assessment of polyploid giant cancer cells and mitotic figures in liver cancer.

Abstract

Primary liver cancer is among the most lethal malignancies, with cell-level structural features such as polyploid giant cancer cells and mitotic figures strongly associated with poor patient prognosis. However, the quantification of these features is hindered by a shortage of pathologists, high workloads, and subjective discrepancies. To address these challenges, we leverage deep learning algorithms to enable the rapid detection of cell-level features, combining this capability with survival analysis to establish a novel, practical prognostic risk assessment system for liver cancer diagnosis and treatment. In collaboration with Peking University Shenzhen Hospital, we collected 172 liver cancer cases, comprising 340 pathology images, to construct the HCCP&M dataset. Our full-process calculation system integrates cell-level feature detection and survival analysis. During the detection phase, the CellFDet framework achieves F1 scores of 0.814, 0.819, and 0.935 for detecting polyploid giant cancer cells, mitotic figures, and general cells, respectively. In the survival analysis phase, patients were stratified into high-risk and low-risk groups based on the polyploid giant cancer cell index (P < 0.0001) and the mitotic index (P = 0.0025), with both indices demonstrating significant survival differences. Correlation analysis further confirmed these features as independent prognostic indicators for liver cancer. Our proposed system not only enables accurate detection of cell-level structural features but also provides reliable survival predictions, offering a valuable tool for improving the prognosis and treatment planning for liver cancer patients.