Application of a deep learning algorithm for the diagnosis of HCC

Abstract

Background & aims: Hepatocellular carcinoma (HCC) is characterized by a high mortality rate. The Liver Imaging Reporting and Data System (LI-RADS) results in a considerable number of indeterminate observations, rendering an accurate diagnosis difficult.

Methods: We developed four deep learning models for diagnosing HCC on computed tomography (CT) via a training-validation-testing approach. Thin-slice triphasic CT liver images and relevant clinical information were collected and processed for deep learning. HCC was diagnosed and verified via a 12-month clinical composite reference standard. CT observations among at-risk patients were annotated using LI-RADS. Diagnostic performance was assessed by internal validation and independent external testing. We conducted sensitivity analyses of different subgroups, deep learning explainability evaluation, and misclassification analysis.

Results: From 2,832 patients and 4,305 CT observations, the best-performing model was Spatio-Temporal 3D Convolution Network (ST3DCN), achieving area under receiver-operating-characteristic curves (AUCs) of 0.919 (95% CI, 0.903-0.935) and 0.901 (95% CI, 0.879-0.924) at the observation (n = 1,077) and patient (n = 685) levels, respectively during internal validation, compared with 0.839 (95% CI, 0.814-0.864) and 0.822 (95% CI, 0.790-0.853), respectively for standard of care radiological interpretation. The negative predictive values of ST3DCN were 0.966 (95% CI, 0.954-0.979) and 0.951 (95% CI, 0.931-0.971), respectively. The observation-level AUCs among at-risk patients, 2-5-cm observations, and singular portovenous phase analysis of ST3DCN were 0.899 (95% CI, 0.874-0.924), 0.872 (95% CI, 0.838-0.909) and 0.912 (95% CI, 0.895-0.929), respectively. In external testing (551/717 patients/observations), the AUC of ST3DCN was 0.901 (95% CI, 0.877-0.924), which was non-inferior to radiological interpretation (AUC 0.900; 95% CI, 0.877--923).

Conclusions: ST3DCN achieved strong, robust performance for accurate HCC diagnosis on CT. Thus, deep learning can expedite and improve the process of diagnosing HCC.

Impact and implications: The clinical applicability of deep learning in HCC diagnosis is potentially huge, especially considering the expected increase in the incidence and mortality of HCC worldwide. Early diagnosis through deep learning can lead to earlier definitive management, particularly for at-risk patients. The model can be broadly deployed for patients undergoing a triphasic contrast CT scan of the liver to reduce the currently high mortality rate of HCC.

Keywords: AI; CT; HCC; Imaging; LIRADS; Liver cancer.

Graphical abstract

Fig. 1

Overall framework of the developed deep learning model, Spatio-Temporal 3D Convolution Network (ST3DCN), for predicting hepatocellular carcinoma (HCC)/non-HCC. (A) Complete architecture of ST3DCN. (B) Details of the efficient channel attention (ECA) and spatial attention (SA) blocks. P3D, pseudo 3D.

Fig. 2

Patient selection process. ¹Relevant clinical data were also collected. ²Post-treatment scans after percutaneous, transcatheter, or radiation therapy are assessed differently via LIRADS and were excluded. ³Two scans had prior intra-abdominal vascular coils inserted, hindering image quality. CT, computed tomography; LI-RADS, Liver Imaging Reporting and Data System.

Fig. 3

AUC curves of different 3D deep learning models. (A) Internal validation for overall patients; (B) internal validation for at-risk patients; (C) external testing for overall patients; (D) external testing for at-risk patients. Among all models, Spatio-Temporal 3D Convolution Network (ST3DCN; blue line) had the highest diagnostic performance, achieving AUCs of 0.862 and 0.919 at the observation and patient level, respectively during internal validation and 0.881 and 0.901, respectively during external testing. (E) When discounting cholangiocarcinomas and liver metastases for overall patients in external testing, AUCs of 0.981 and 0.982, respectively were achieved. 3DResNet, Three-Dimensional Residual Network; 3DSE, Three-Dimensional Squeeze-and-Excitation; C3D, Convolutional Three-Dimensional.

Fig. 4

Visualization of deep learning classification results. Heatmap plots for three different slices from the same computed tomography (CT) scan of three patients (A) with hepatocellular carcinoma (HCC) and (B) non-HCC. The odd rows depict the CT scan in the portovenous phase. For heatmaps in the even rows, the red color indicates the most risky area for HCC, while the blue area indicates the least risky area. Grad-CAM, gradient-weighted Class Activation Mapping.