IDDF2020-ABS-0078 Immunoscore Classification from Hepatocellular Carcinoma Histopathology Images Using Deep Learning: A Preliminary Study
Qiaofeng Chen, Yunquan Gu, Ruixuan Wang, Xiao Han, Sui Peng, Ming Kuang · Abstracts · 2020
Background Immunotherapy is a recent advance for the treatment of hepatocellular carcinoma (HCC). Immunoscore assessment plays a critical role in precision immunotherapy and can predict prognosis in patients with HCC. This study aims to develop a deep-learning model to automated analyze histopathology images for classification of immunoscore (CD3 or CD8, 0–2 vs. 3–4) in HCC. Methods We trained a patch-based deep convolutional neural network (Resnet-18) on whole-slide images to automatically classify immunoscore into 0–2 or 3–4. The data were randomly split into a training and testing dataset. The performance was first estimated on the training dataset with nine-folded cross-validation and then further validated on the testing dataset. Cross-entropy was used as a model-optimized loss function and the accuracy as well as the area under the receiver operating characteristic curve (AUC) were calculated for the identification values. Heatmaps were also generated by our model to visualize the regions the most associated with the classification. Results We included 28 images from a study cohort of 28 HCC patients for training (18 images) and testing (10 images) the model. After iterative training, an optimized architecture achieved an AUC of 0.71 was used as our final model. For validation on the testing dataset, the model yielded an accuracy of 90% and AUC of 0.93 (95% CI: 0.76 to 1.00) while the percentage of patches positively classified, and outperforms average of the probabilities of the corresponding patches (accuracy 70%; AUC 0.79, 95% CI: 0.50 to 1.00) using the same optimal threshold of 0.33. The heatmaps show that almost all of patches are highly identified to show the regions of immunoscore ((figure 1) A. Immunoscore of 3–4 [positive]. B. Immunoscore of 0–2 [negative]). Conclusions The automated deep-learning model achieved good performance and could potentially assist clinicians in the identification of HCC patients who are more likely to respond to immunotherapy, or at least, providing second opinions on therapeutic decision-making.