Enhancing noncontrast computed tomography for the diagnosis of hepatocellular carcinoma via novel image-guided deep learning artificial intelligence algorithm

Abstract

Background: Liver image classification via Liver Imaging Reporting and Data System (LI‐RADS) results in considerable proportions of indeterminate observations, delaying any definite diagnosis of hepatocellular carcinoma (HCC). Our developed Multiscale Three‐Dimensional Convolutional Network (MS3DCN) deep learning model can achieve a high diagnostic performance on computed tomography (CT). The performance of noncontrast CT enhanced by deep learning models in diagnosing HCC remains uninvestigated. Methods: We collected thin‐cut (<1.25 mm) tri‐phasic CT liver scans and relevant clinical information from six medical centers, with diagnosis of HCC defined via American Association for the Study of Liver Disease guidelines, and confirmed via clinical composite reference standard based on subsequent 12‐month follow‐up. All CT observations were annotated following LI‐RADS by a specialist radiologist blinded to corresponding clinical information. After randomly splitting in a 7:3 ratio to internal training and validation cohorts, we applied MS3DCN (Figure 1A) to all liver observations, with noncontrast, arterial and portovenous phases validated separately. We additionally developed, trained and validated a novel deep learning model, Convolutional Block Attention Module (CBAM, Figure 1B), based solely on noncontrast CT images. Both models were further externally tested on archived thick‐cut (≥5 mm) images reconstructed with a generative adversarial network‐based lossless image compression model. Results: Among 2281 patients with thin‐cut scans, 686 (30.1%) were HCC. CBAM achieved an area under curve (AUC) of 0.844 (95%CI 0.813‐0.874), positive predictive value (PPV) of 0.782 (95%CI 0.705‐0.859) and negative predictive value of 0.779 (95%CI 0.745‐0.813), numerically better than MS3DCN (AUC 0.814 [95%CI 0.780‐0.848], PPV 0.631 [95%CI 0.546‐0.717] and NPV 0.975 [95%CI 0.947‐0.994]), and comparable to LI‐RADS (AUC 0.852 [95%CI 0.820‐0.881], PPV 0.975 [95%CI 0.947‐0.994], NPV 0.883 [95%CI 0.885‐0.909]). AUCs of MS3DCN for arterial and portovenous phases were 0.967 (95%CI 0.946‐0.980) and 0.969 (95%CI 0.954‐0.981) respectively. Among 432 patients with archived thick‐cut scans (46.8% HCC), CBAM's performance remained suboptimal, with an AUC of 0.661 (95%CI 0.568‐0.754). Conclusion: After deep learning enhancement, noncontrast thin‐cut CT’s diagnostic performance for HCC was non‐inferior to radiologist reporting via LI‐RADS. Purposely‐developed models based on noncontrast image architecture performed better than models designed for multi‐phasic CT. Potential future applications of artificial intelligence‐enhanced noncontrast CT include opportunistic screening, automatic reporting, and HCC surveillance in at‐risk populations if radiation risks can be dampened.