Investigation of deep learning model for predicting immune checkpoint inhibitor treatment efficacy on contrast-enhanced computed tomography images of hepatocellular carcinoma

Although the use of immune checkpoint inhibitors (ICIs)-targeted agents for unresectable hepatocellular carcinoma (HCC) is promising, individual response variability exists. Therefore, we developed an artificial intelligence (AI)-based model to predict treatment efficacy using pre-ICIs contrast-enhanced computed tomography (CT) imaging characteristics. We evaluated the efficacy of atezolizumab and bevacizumab in 43 patients at the Nagasaki University Hospital from 2020 to 2022 using the modified Response Evaluation Criteria in Solid Tumors. A total of 197 Progressive Disease (PD), 271 Partial Response (PR), and 342 Stable Disease (SD) contrast CT images of HCC were used for training. We used ResNet-18 as the Convolutional Neural Network (CNN) model and YOLOv5, YOLOv7, YOLOv8 as the You Only Look Once (YOLO) model with precision-recall curves and class activation maps (CAMs) for diagnostic performance evaluation and model interpretation, respectively. The 3D t-distributed Stochastic Neighbor Embedding was used for image feature analysis. The YOLOv7 model demonstrated Precision 53.7%, Recall 100%, F1 score 69.8%, mAP@0.5 99.5% for PD, providing accurate and clinically versatile predictions by identifying decisive points. The ResNet-18 model had Precision 100% and Recall 100% for PD. However, the CAMs sites did not align with the tumors, suggesting the CNN model is not predicting that a given CT slice is PD, PR, or SD, but that it accurately predicts Individual Patient's CT slices. Preparing substantial training data for tumor drug effect prediction models is challenging compared to general tumor diagnosis models; hence, large-scale validation using an efficient YOLO model is warranted.