Author: Enhui Chang, Yunfei Dong, Yifei Hao, Chengliang Jin, Shengsheng Lai, Yi Long, Mengni Wu, Yulu Wu, Ruimeng Yang, Zhenyu Yang, Yue Yuan, Lei Zhang, Wanli Zhang, Yaogong Zhang 👨🔬
Affiliation: Duke Kunshan University, Department of Radiology, The Second Affiliated Hospital, School of Medicine, South China University of Technology, Medical Physics Graduate Program, Duke Kunshan University 🌍
Purpose: Macrotrabecular-Massive Hepatocellular Carcinoma (MTM-HCC) is one type of liver cancer showed minimum image signature for accurate non-invasive diagnosis. This study aims to develop and evaluate novel deep learning models in the classification of MTM-HCC preoperatively using CT images from multiple centers.
Methods: A total of 538 patients (373 hospital A: 87 MTM, 286 non-MTM; 165 hospital B: 43 MTM, 122 non-MTM) were recruited from two institutions. Two-dimensional CT images were extracted and enhanced by contrast limited adaptive histogram equalization (CLAHE). CT images were segmented, and VOIs were manually delineated by experienced radiologists. Data from Hospital A and Hospital B were utilized as the training set and testing set, respectively. A weighted loss function (WLF) was used to address data imbalance issue. Multiple network architectures were explored, including CNN, ResNet50, ResNet101, and Graph Neural Networks (GNN), each combined with Long Short-Term Memory (LSTM) which was initially designed to include temporal dependency information. Notably, we innovatively utilized GNN to capture spatial relationships and features within 2D images, and LSTM to capture inter-slice spatial dependencies. Model performance was assessed using accuracy as the primary evaluation metric.
Results: The top-performing model, GNN+LSTM, reached 78.18% of accuracy, outperformed ResNet-LSTM models (75.15%–75.76%), ResNet101 (73.94%), ResNet50 (73.94%), and CNN (70.19%). It seems that by incorporating additional spatial information in 2D images by GNN model, and inter-slice spatial dependencies by LSTM model, the highest accuracy in the MTM-HCC diagnosis was achieved in this study.
Conclusion: This work investigated and showed the advantage of a novel GNN-LSTM model in the preoperative MTM-HCC diagnosis using single-modality, multi-center CT images. Notably, the performance of the proposed GNN-LSTM model, despite based on single-modality CT images, approaches the performance of recent multi-model studies. Future work will incorporate additional clinical data and multi-modal information to further boost the non-invasive diagnostic accuracy of MTM-HCC.