Primary liver cancer classification from routine tumour biopsy using weakly supervised deep learning.

Area of Science:

• Digital pathology
• Machine learning in oncology
• Cancer diagnostics

Background:

• Diagnosing primary liver cancers (PLCs), particularly combined hepatocellular-cholangiocarcinoma (cHCC-CCA), on biopsies presents significant challenges.
• Routine H&E-stained biopsies are standard for pathological examination.

Purpose of the Study:

• To develop and validate a weakly supervised learning method for automatic classification of PLCs on routine biopsies.
• To assess the model's ability to differentiate between hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (iCCA), and aid in cHCC-CCA identification.

Main Methods:

• A ResNet18 neural network was trained on 166 PLC biopsies using weakly supervised learning with tumour/non-tumour annotations.
• Unsupervised clustering was applied to extracted tile features without prior knowledge of malignancy labels.
• The model was validated on internal and external datasets.

Main Results:

• The two-cluster model successfully distinguished between HCC and iCCA histological features.
• High diagnostic agreement was achieved for HCC (100% internal, 96% external) and iCCA (78% internal, 87% external).
• cHCC-CCA cases showed variable proportions of tiles from both clusters, suggesting a potential diagnostic indicator.

Conclusions:

• The weakly supervised learning method can identify specific morphological features of HCC and iCCA on H&E biopsies.
• While not directly identifying cHCC-CCA, the model's ability to quantify HCC and iCCA tile proportions can assist in diagnosing this challenging subtype.