Annals of surgical oncology · 2026
Foundation Model and Multi-Instance Learning-Based Framework for Predicting Lymphovascular Invasion in Prostate Cancer.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Zheng Qingyuan, Mei Haonan, Wang Dan, Liu Xiuheng, Wang Lei, Chen Zhiyuan
Аннотация
BACKGROUND: Lymphovascular invasion (LVI) is a well-established adverse prognostic factor in prostate cancer (PCa). This study aimed to develop and validate an artificial intelligence (AI)-based framework leveraging multi-instance learning (MIL) and foundation models for accurate and interpretable prediction of LVI in prostate cancer using whole-slide images (WSIs).
METHODS: A weakly supervised deep-learning pipeline based on the clustering-constrained attention MIL framework was implemented to analyze hematoxylin and eosin (H&E)-stained WSIs from two independent cohorts: 280 patients from Renmin Hospital of Wuhan University (RHWU) and 340 patients from The Cancer Genome Atlas (TCGA). Feature extraction was performed using pretrained encoders including UNI-v2, CONCH, and ResNet-50. Attention heatmaps were used to interpret model focus, whereas biologic correlates of model predictions were explored through differential expression analysis and gene ontology (GO) enrichment.
RESULTS: The proposed models achieved strong predictive performance, with UNI-v2 outperforming the other encoders (area under the curve [AUC], 0.839 for RHWU and 0.854 for TCGA). Attention-based interpretability highlighted high-risk histopathologic regions characterized by hyperchromatic nuclei, prominent nucleoli, and increased mitotic activity. Exploratory transcriptomic analysis showed 381 differentially expressed genes (DEGs) between LVI-positive and LVI-negative groups. Gene ontology enrichment showed that upregulated DEGs in the LVI-positive group were enriched in mitotic and immune-related pathways, whereas downregulated genes were associated with ion transport.
CONCLUSIONS: This study exhibited a robust and interpretable AI framework for predicting LVI in PCa from WSIs using weakly supervised learning and domain-adapted foundation models. The model achieved high accuracy, provided biologically meaningful insights, and showed potential for clinical translation as a decision-support tool in precision pathology.
METHODS: A weakly supervised deep-learning pipeline based on the clustering-constrained attention MIL framework was implemented to analyze hematoxylin and eosin (H&E)-stained WSIs from two independent cohorts: 280 patients from Renmin Hospital of Wuhan University (RHWU) and 340 patients from The Cancer Genome Atlas (TCGA). Feature extraction was performed using pretrained encoders including UNI-v2, CONCH, and ResNet-50. Attention heatmaps were used to interpret model focus, whereas biologic correlates of model predictions were explored through differential expression analysis and gene ontology (GO) enrichment.
RESULTS: The proposed models achieved strong predictive performance, with UNI-v2 outperforming the other encoders (area under the curve [AUC], 0.839 for RHWU and 0.854 for TCGA). Attention-based interpretability highlighted high-risk histopathologic regions characterized by hyperchromatic nuclei, prominent nucleoli, and increased mitotic activity. Exploratory transcriptomic analysis showed 381 differentially expressed genes (DEGs) between LVI-positive and LVI-negative groups. Gene ontology enrichment showed that upregulated DEGs in the LVI-positive group were enriched in mitotic and immune-related pathways, whereas downregulated genes were associated with ion transport.
CONCLUSIONS: This study exhibited a robust and interpretable AI framework for predicting LVI in PCa from WSIs using weakly supervised learning and domain-adapted foundation models. The model achieved high accuracy, provided biologically meaningful insights, and showed potential for clinical translation as a decision-support tool in precision pathology.