Predicting Survival After Hepatocellular Carcinoma Resection Using Deep Learning on Histological Slides

Area of Science:

• Digital pathology
• Artificial intelligence in oncology
• Hepatocellular carcinoma research

Background:

• Standardized risk stratification is crucial for hepatocellular carcinoma (HCC) treatment strategies.
• Accurate prognosis is needed to evaluate adjuvant therapies post-resection/ablation for HCC.

Purpose of the Study:

• To develop and validate deep-learning algorithms for predicting survival in HCC patients undergoing surgical resection.
• To compare the performance of AI models against traditional prognostic scores.

Main Methods:

• Two deep-learning algorithms (SCHMOWDER and CHOWDER) were developed using whole-slide imaging (WSI) of digitized histological slides.
• Models were trained on a discovery set (n=194) and validated on an independent set (TCGA, n=328).
• Feature extraction involved dividing WSIs into tiles and using a pretrained convolutional neural network; SCHMOWDER incorporated pathologist-annotated tumoral areas.

Main Results:

• Deep-learning models achieved high c-indices for survival prediction (SCHMOWDER: 0.78, CHOWDER: 0.75) in the discovery set.
• Both AI models demonstrated superior predictive performance compared to a composite score of baseline variables in both discovery and validation sets.
• Pathological analysis identified vascular spaces, macrotrabecular pattern, and lack of immune infiltration as key indicators of poor survival.

Conclusions:

• Artificial intelligence significantly enhances the prediction of HCC prognosis.
• The integration of pathologist expertise with machine learning algorithms improves AI model development and biological interpretability.
• AI-driven risk stratification can refine therapeutic strategies for HCC patients.