Background: Hepatocellular carcinoma (HCC) is the third leading cause of cancer-related deaths worldwide. Worryingly, the rate of recurrence among patients who have undergone curative treatment can be as high as 88%. Despite recent technological advancements in the field, tumor recurrence remains a significant challenge, necessitating careful reappraisal of patient and disease status. To address this, we aimed to develop a precise deep learning algorithm to predict liver cancer recurrence utilizing whole slide images (WSIs). Methods: We developed an attention-based deep learning model to predict liver cancer recurrence from digital slides of hematoxylin and eosin-stained liver tissue. The dataset, sourced from the TCGA database, underwent preprocessing that included tissue segmentation, tiling, and extraction of histopathological features. The labeled dataset was randomly split into training (70%), validation (15%), and testing (15%) sets. The attention-based model was trained on the training set and optimized using the validation set, while the testing set was kept completely unseen to be used only for the final evaluation of the model's performance. Performance was evaluated using standard metrics like AUC, sensitivity, specificity, and accuracy. Attention heatmaps were generated to provide interpretability and insights into the model's decision-making process. Additionally, we trained a Hover-Net model to analyze the distribution and organization of cells in the liver cancer microenvironment, comparing recurrent and non-recurrent cases. The study is supported by the NIH-NCI for Next Generation Onco-Pathologists Program at our institution. Results: The dataset comprised 450 whole slide images (WSIs), which were classified as either post-therapeutic liver cancer recurrence or no recurrence. The model demonstrated robust performance on the validation dataset, achieving an Area Under the Curve (AUC) of ~0.70. Conclusion: Our study presents an innovative deep learning approach that accurately predicts liver cancer recurrence utilizing H&E whole slide images (WSIs). By leveraging attention-guided deep neural networks, we were able to develop a powerful prognostic tool for forecasting the risk of liver cancer recurrence. These findings have critical implications for optimizing personalized therapeutic interventions and surveillance strategies in liver cancer management, thus advancing the field of precision pathology. Additional, experiments are being conducted to further expand our findings.

Learning Objectives 

1. Provide a fundamental understanding of Convolutional Neural Networks (CNNs) and Multiple Instance Learning (MIL) for extracting meaningful histological features and unlocking novel histopathological patterns.              
2. Showcase real-world applications and case studies (n=450) that demonstrate the benefits of MIL in the stratification of liver cancer recurrence.           
3. Understand the pivotal role of pathologists in utilizing AI-powered risk assessment to integrate precision diagnostics, transform workflows, and manage population health.