Detection of hepatocellular carcinoma with noncontrast computed tomography and artificial intelligence

Abstract

Liver cancer poses a heavy burden of disease worldwide, being the sixth most common and third most deadly, causing over 90,000 incident cases and 83,000 deaths annually. Hepatocellular carcinoma (HCC) is the commonest form of liver cancer, with its high mortality rate and poor prognosis partly due to its insidious clinical course and advanced stage at diagnosis, highlighting the importance of early detection and early treatment. Currently, HCC is often diagnosed radiologically, with contrast-enhanced computed tomography (CT) or magnetic resonance imaging (MRI), using the Liver Imaging Reporting and Data System (LI-RADS). However, LI-RADS has several limitations – most importantly, it results in a large number of intermediate observations with no recommended definitive management, potentially leading to significant diagnostic delays. The potential value of non-contrast scans in the opportunistic screening of HCC has often been overlooked. Artificial intelligence (AI) and deep learning (DL) may be capable of detecting subtle changes in non-contrast scans to detect HCC. While many studies have investigated the use of AI for HCC diagnosis on contrast-enhanced CT, there is a lack of research into the potential of AI and DL for HCC detection in non-contrast scans. In our study, we developed, trained and evaluated a three-dimensional Convolutional Block Attention Module (CBAM) model for the detection of HCC on non-contrast CT scans. Thincut triphasic CT scans were collected from five medical centers in our locality, and randomly divided by a 7:3 ratio into two datasets for training and internal validation. Diagnosis of HCC followed American Association for the Study of Liver Disease guidelines, and was confirmed via clinical composite reference standard based on subsequent 12-month follow-up. The model was evaluated on the internal validation cohort as well an independent external testing cohort. Sensitivity analyses were performed on several patient subgroups to evaluate model robustness and generalizability. In order to improve model explainability, we also generated heatmaps to visualize the model’s decision-making process. 2,223 patients were included in the internal cohort, among which 675 (30.4%) were diagnosed to have HCC. In the internal validation cohort, the CBAM model achieved an area under curve (AUC) of 0.807 (95%CI 0.772-0.841), comparable to that of radiological interpretation at 0.851 (95%CI 0.820-0.882). Among at-risk patients, definite HCC cases, indeterminate scans, and scans with small lesions < 2 cm in size, the model attained AUCs of 0.769 (95%CI 0.721-0.817), 0.815 (95%CI 0.778-0.853), 0.769 (95%CI 0.704-0.834) and 0.773 (95%CI 0.692-0.854), which were all non-inferior to radiological interpretation. In the external cohort containing 584 patients, the CBAM achieved an AUC of 0.789 (95%CI 0.750-0.827), as compared to 0.927 (95%CI 0.904-0.950) for radiological interpretation. In conclusion, the CBAM model demonstrated robust diagnostic capabilities, achieving a diagnostic performance comparable to that of radiological interpretation in the internal dataset as well as all analyzed subgroups. This shows that non-contrast CT combined with AI is capable of serving as an efficient and accurate method for HCC detection, and may also have potential future applications in opportunistic screening for HCC.