Novel deep learning radiomics model for preoperative evaluation of hepatocellular carcinoma differentiation based on computed tomography data

Abstract

The evaluation of tumor differentiation is an urgent clinical issue that would facilitate the establishment of individualized therapeutic strategies.1-3 Our team developed a deep learning radiomics model based on computed tomography (CT) data for preoperative evaluation of hepatocellular carcinoma (HCC) differentiation (low vs. high grade) and preliminarily explored the biological basis of the radiomics model. We included 1047 patients from the First Affiliated Hospital, College of Medicine, Zhejiang University (Institution 1) and 187 patients from the Ningbo Medical Center Lihuili Hospital (Institution 2). Data from Institution 1 were divided into training and internal validation cohorts by stratified sampling at a 3:1 ratio, while data from Institution 2 constituted the independent test cohort (Figure S1). Patient characteristics are shown in Table 1; there were no significant differences in the distribution of clinical characteristics among the three cohorts. The radiomics pipeline (Figure 1) mainly involved data acquisition from CT images (Method S1), segmentation of regions of interest, feature extraction (Table S1) and selection, model construction and evaluation and multiomics analysis (Method S2). In total, 707 radiomics features were extracted from CT image data; 614 were filtered out because of low reproducibility or high redundancy, and 25 features with a significant impact on the target were ultimately selected (Table S2). A radiomics signature was established using the random forest (RF) method (Table S3, Figure S2). The AUCs in the training, internal validation and external test cohorts were 0.82, 0.76 and 0.75, respectively (Figure S3). Violin plots of selected features are shown in Figure 4A. The accuracy of the radiomics signature in the training, validation and test cohorts were 0.75, 0.72, and 0.66, respectively; the sensitivity was 0.76, 0.70, and 0.74, respectively; and the specificity was 0.72, 0.75, and 0.54, respectively. The deep learning model in this study was modified from VGG194 (Table S4). A illustration of deep learning model structure is shown in Figure 2. The AUCs of the deep learning model in the training, internal validation and test cohorts were 0.85, 0.81, and 0.75, respectively (Figure S4). The model had an accuracy of 0.77, 0.75, and 0.66, respectively; sensitivity of 0.76, 0.81, and 0.62, respectively; and specificity of 0.66, 0.66, and 0.72, respectively in the three cohorts. In the comparison of the deep learning model with the radiomics signature, p values from the DeLong test5 were 0.09, 0.17, and 0.62 in the training, validation, and test cohorts, respectively. There were no significant differences between the deep learning model and radiomics signature, although the former had a slightly higher AUC. To see how much value radiomics or deep learning can bring to some risk factors about tumor morphology and size, the features (original_shape2D_Sphericity, original_shape2D_Elongation, original_shape2D_MajorAxisLength) were used to construct a morphological model (Figure S5). Quantitative indices in the comparisons between the clinical model, radiomics signature, deep learning model, and fused model and the results of the DeLong test are summarized in Table 2. The fused model showed the best performance in the training, validation, and test cohorts, with an AUC of 0.89, 0.83, and 0.80, respectively; accuracy of 0.82, 0.77, and 0.73, respectively; sensitivity of 0.85, 0.81, and 0.71, respectively; specificity of 0.76, 0.71, and 0.75, respectively; PPV of 0.84, 0.80, and 0.79, respectively; NPV of 0.78, 0.73, and 0.66, respectively; and F1 score of 0.77, 0.72, and 0.71 respectively. The calibration curves showed that the fused model had better concordance between predicted and actual probabilities than the other models (Figure 3D). Comparison of the decision curves of the four models in the test set indicated that the fused model had greater clinical utility (Figure 3E), and the IDI indicated that the predicted probabilities of the fused model were significantly improved compared to those of the other models (Figure S7). A nomogram for preoperative prediction of HCC pathologic grade was established based on the fused model (Figure 3F). A total of 69 patients with CT data were included in the multiomics analysis. After data preprocessing, 19723 genomics, 42807 transcriptomics, and 3658 proteomics variables with differential expression between high- and low-grade HCC (valid data > 80%) were extracted. Pearson's correlation coefficients between radiomics features and multiomics variables are shown as correlation heat maps (Figure 4A). The selected radiomics features reconstructed 65.54%, 64.65%, and 72.69% of the differentially expressed genes, transcripts, and proteins (Figure 4B). The coverage of each type of -omics was 60% with just 15 radiomics features. The radiomics-related multiomics variables showed significant differences between the different pathologic grades (high vs. low grade) (Figure 4C). The results of the gene enrichment analysis of 25 radiomics features are summarized in Figure 4D. In the enrichment result for wavelet_LL_first-order_entropy, 21 GO terms and pathways were identified that are potentially related to HCC development. For example, wavelet_LL_first-order_entropy was associated with abnormal alcohol dehydrogenase activity, which leads to abnormal development and cell apoptosis. Key genes associated with original_shape2D_sphericity were related to the phosphatidylinositol 3-kinase (PI3K)/protein kinase B (AKT) signaling pathway (Figure 4F), which is involved in apoptosis, cancer cell proliferation, DNA repair, and cancer differentiation, among other biological processes. In conclusion, we established a deep learning radiomics model that can be used for preoperative pathological grading of HCC and served as a noninvasive prediction tool to guide clinical decision-making. We would like to thank the patients who participated in this study. This work was supported by the National Key Research and Development Program of China (grant number: 2018YFE0183900), the Natural Science Foundation of China (NSFC grant number: 81971686) and the Scientific Research Fund of Zhejiang Provincial Education Department (grant number: Y202045565). The authors declare that they have no competing interests. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.

Topics & Concepts