Deep learning for automatic detection of hepatocellular carcinoma in dynamic contrast-enhanced MRI.

[OBJECTIVE] Develop a deep learning model for automatic hepatocellular carcinoma (HCC) detection in T1 weighted imaging (WI) Dynamic Contrast-Enhanced (DCE) liver MRI using extracellular contrast agent, and to analyze its performance at both patient and lesion levels.

[MATERIALS AND METHODS] This retrospective study included two cohorts, the first included patients (N = 296) undergoing HCC surveillance with diagnosed HCC as well as negative cases. The 233 HCC negative patients and 12 HCC positive patients were used to create the HCC Surveillance test set, aiming to evaluate patient level performance on simulated screening conditions. The second included Pre-Ablation patients (N = 67), all positive for HCC, used as test set for lesion level evaluation and to measure generalization performance. The two largest public liver lesion datasets (CT, N = 1037 and MRI, N = 485) were used for pre-training the algorithms. An attention U-Net model was trained to segment and detect HCC and was compared to the state-of-art nnU-Netv2. Diagnostic accuracy was evaluated using sensitivity, specificity, mean false positives per patient, PPV and NPV, the Area Under the Curve (AUC) of the Free-Response Operating Characteristic (FROC) curves and the Receiver Operating Characteristic (ROC) curves.

[RESULTS] The final population included 363 patients (58 ± 11 years; 284 men; 247 lesions): 51 HCC positive patients (113 lesions) used in training set, 245 patients (12 HCC positive with 21 lesions, 233 HCC negative) in the HCC Surveillance testing set, 67 HCC positive patients (113 lesions) in the HCC Pre-Ablation testing set. At patient level, 83% sensitivity and 72% specificity [AUC of 0.80 (95% CI: 0.66-0.91)] was measured on the HCC Surveillance test set. At lesion level, 80% of sensitivity for a mean false positive per patient of 1 was measured on the HCC Pre-Ablation test set with the pre-trained model with a FROC AUC of 0.82 (95% CI: 0.77-0.88), significantly outperforming the nnU-Netv2 at 0.61 (95% CI: 0.52-0.69, p < 0.01).

[CONCLUSIONS] Both patient-level and lesion-level achieved 80% HCC detection sensitivity by using a deep learning segmentation neural network pre-trained from large open datasets. This performance highlights the translational potential of such tools in the clinical workup of patients at risk of HCC.