SiameseNet based on multiple instance learning for accurate identification of the histological grade of ICC tumors

TL;DR

A new deep learning model called SiameseNet improves the accuracy of predicting the histological grade of intrahepatic cholangiocarcinoma tumors.

Contribution

The novel SiameseNet framework uses multiple instance learning and cross-attention to address tumor heterogeneity in ICC grade prediction.

Findings

The model achieved 86.0% accuracy and 86.2% AUC in predicting ICC histological grade.

Cross-attention mechanisms improved feature representation and model robustness.

The framework shows potential for clinical use in ICC histopathological assessment.

Abstract

After hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma (ICC) is the second most common primary liver cancer. Timely and accurate identification of ICC histological grade is critical for guiding clinical diagnosis and treatment planning.

We proposed a dual-branch deep neural network (SiameseNet) based on multiple-instance learning and cross-attention mechanisms to address tumor heterogeneity in ICC histological grade prediction. The study included 424 ICC patients (381 in training, 43 in testing). The model integrated imaging data from two modalities through cross-attention, optimizing feature representation for grade classification.

In the testing cohort, the model achieved an accuracy of 86.0%, AUC of 86.2%, sensitivity of 84.6%, and specificity of 86.7%, demonstrating robust predictive performance.

The proposed framework effectively mitigates performance degradation caused by tumor heterogeneity. Its high accuracy and generalizability suggest potential clinical utility in assisting histopathological assessment and personalized treatment planning for ICC patients.

Figures