A multi‐scale, multi‐region and attention mechanism‐based deep learning framework for prediction of grading in hepatocellular carcinoma. Issue 4 (10th December 2022)

• Record Type: Journal Article
• Authors: Wei, Jingwei; Ji, Qian; Gao, Yu; Yang, Xiaozhen; Guo, Donghui; Gu, Dongsheng; Yuan, Chunwang; Tian, Jie; Ding, Dawei
• Is Part Of: Medical physics. Volume 50:Issue 4(2023)
• Page Start: 2290
• Page End: 2302
• Publication Date: 2022-12-10
• Subjects: computed tomography -- deep learning -- hepatocellular carcinoma -- histopathological grading -- radiomics; Medical physics -- Periodicals; Medical physics; Geneeskunde; Natuurkunde; Toepassingen; Biophysics; Periodicals; Periodicals; Electronic journals; 610.153
• DOI: 10.1002/mp.16127
• Languages: English
• ISSNs: 0094-2405

Abstract

Abstract: Background: Histopathological grading is a significant risk factor for postsurgical recurrence in hepatocellular carcinoma (HCC). Preoperative knowledge of histopathological grading could provide instructive guidance for individualized treatment decision‐making in HCC management. Purpose: This study aims to develop and validate a newly proposed deep learning model to predict histopathological grading in HCC with improved accuracy. Methods: In this dual‐centre study, we retrospectively enrolled 384 HCC patients with complete clinical, pathological and radiological data. Aiming to synthesize radiological information derived from both tumour parenchyma and peritumoral microenvironment regions, a modelling strategy based on a multi‐scale and multi‐region dense connected convolutional neural network (MSMR‐DenseCNNs) was proposed to predict histopathological grading using preoperative contrast enhanced computed tomography (CT) images. Multi‐scale inputs were defined as three‐scale enlargement of an original minimum bounding box in width and height by given pixels, which correspondingly contained more peritumoral analysis areas with the enlargement. Multi‐region inputs were defined as three regions of interest (ROIs) including a squared ROI, a precisely delineated tumour ROI, and a peritumoral tissue ROI. The DenseCNN structure was designed to consist of a shallow feature extraction layer, dense block module, and transition and attention module. The proposed MSMR‐DenseCNN was pretrained by the ImageNet dataset to capture basic graphic characteristics from the images and was retrained by the collected retrospective CT images. The predictive ability of the MSMR‐DenseCNN models on triphasic images was compared with a conventional radiomics model, radiological model and clinical model. Results: MSMR‐DenseCNN applied to the delayed phase (DP) achieved the highest area under the curve (AUC) of 0.867 in the validation cohort for grading prediction, outperforming those on the arterial phase (AP) and portal venous phase (PVP). Fusion of the results on triphasic images did not increase the predictive ability, which underscored the role of DP for grading prediction. Compared with a single‐scale and single‐region network, the DP‐phase based MSMR‐DenseCNN model remarkably raised sensitivity from 67.4% to 75.5% with comparable specificity of 78.6%. MSMR‐DenseCNN on DP defeated conventional radiomics, radiological and clinical models, where the AUCs were correspondingly 0.765, 0.695 and 0.612 in the validation cohort. Conclusions: The MSMR‐DenseCNN modelling strategy increased the accuracy for preoperative prediction of grading in HCC, and enlightens similar radiological analysis pipelines in a variety of clinical scenarios in HCC management.