Deep Learning Classification and Quantification of Pejorative and Nonpejorative Architectures in Resected Hepatocellular Carcinoma from Digital Histopathologic Images

Astrid Laurent-Bellue ¹, Aymen Sadraoui ², Laura Claude ³

⁰Department of Pathology, Bicêtre Hospital, Assistance Publique-Hôpitaux de Paris, Le Kremlin-Bicêtre, France.

The American Journal of Pathology +

|

June 15, 2024

View abstract on PubMed

Summary

Artificial intelligence can now identify high-risk hepatocellular carcinoma (HCC) recurrence using deep learning on liver tissue images. This approach aids in selecting patients for adjuvant therapy after surgery.

Area Of Science

Hepatobiliary pathology
Computational pathology
Artificial intelligence in medicine

Background

Hepatocellular carcinoma (HCC) recurrence after liver resection is common, necessitating risk stratification for adjuvant therapy.
Accurate histologic predictors of recurrence, such as microvascular invasion (mVI), are time-consuming and imperfect to evaluate.
Identifying patients at high risk of recurrence is crucial for optimizing post-operative treatment strategies.

Purpose Of The Study

To develop and validate a deep learning algorithm for identifying and quantifying adverse histologic architectures associated with HCC recurrence.
To assess the algorithm's performance in predicting recurrence risk and its correlation with established predictors like mVI.
To establish a foundation for an AI-based composite predictive algorithm for early HCC recurrence.

Main Methods

A supervised deep learning model (ResNet34) was trained on 680 whole slide images (WSIs) from 107 liver resection specimens.
The algorithm was designed to identify and quantify pejorative architectures indicative of recurrence risk.
Model performance was evaluated at patch and WSI levels and validated on an external cohort of 29 HCC cases.

Main Results

The AI model achieved high accuracy in identifying pejorative architectures (0.864 patch level, 0.823 WSI level).
External validation demonstrated good generalization capabilities with 0.787 WSI level accuracy.
Quantified pejorative architectures correlated positively with microvascular invasion and tumor emboli, suggesting they are reliable surrogates.

Conclusions

AI-driven identification of pejorative architectures is a promising, efficient surrogate for microvascular invasion in HCC.
This approach offers strong predictive value for early post-resection recurrence risk.
The study represents a significant first step towards an AI-integrated predictive algorithm for HCC recurrence.