[Development and prognostic value of sequential feature reconstruction-based multi-instance learning model for survival prediction of gastric cancer whole-slide images].

Objective: To construct a survival prognostic prediction model for whole-slide histopathological images (WSIs) of gastric cancer based on multi-instance learning with sequential feature reconstruction, and to explore its predictive value. Methods: Retrospectively collected hematoxylin-eosin stained WSIs and clinical prognostic data of gastric cancer patients treated at the 1st, 8th and 9th Medical Centers of Chinese PLA General Hospital from January 2015 to December 2022. The sample sizes of the three medical centers were 386, 284 and 216 cases, denoted as Dataset 1, Dataset 2 and Dataset 3 respectively. This study developed a multi-instance learning model based on sequential feature reconstruction, named sequence reordering and sparse autoencoder-based enhanced representation transformer for multiple instance learning (S²ERT-MIL). After extracting patch features of pathological images via a pre-trained model, the model reconstructed and optimized patch instance features through sequence reordering, regional feature enhancement, sparse autoencoding and cross-regional information fusion, and finally output patient-level survival risk scores. Five-fold cross-validation was adopted to evaluate model performance, and comparisons were conducted with representative MIL models including Attention-Based Multiple Instance Learning (ABMIL), Clustering-constrained Attention Multiple Instance Learning (CLAM), Dual-Scale Multiple Instance Learning (DSMIL), Dual-Tier Feature Distillation Multiple Instance Learning (DTFD-MIL) and Representation Refinement and Transformation Multiple Instance Learning (RRT-MIL). The concordance index (C-index) was used to assess the survival prediction performance of the model. Patients were divided into high-risk and low-risk groups according to the median risk score, and the Kaplan-Meier method with Log-rank test was applied to evaluate the risk stratification ability. Visualization analysis of high- and low-attention patches was also performed. Results: S²ERT-MIL achieved the highest C-index across all three datasets, with values of 0.7493±0.0486, 0.6758±0.0347 and 0.7232±0.0362 in Dataset 1, Dataset 2 and Dataset 3, respectively, all higher than those of the comparison models. Kaplan-Meier survival analysis showed that S²ERT-MIL stratified patients into high-risk and low-risk groups with significantly different survival outcomes in all three datasets. Attention-region visualization showed that high-attention patches were mainly derived from areas with dense tumor cells, complex tissue architecture or marked stromal reaction, whereas low-attention patches mostly represented low-information regions or areas relatively weakly associated with prognosis. Conclusion: S²ERT-MIL can predict patient survival prognosis based on H&E-stained gastric cancer WSIs and showed favorable risk discrimination and stratification performance across three datasets.