Table 1_Deep learning–based auto-segmentation and RECIST evaluation after concurrent chemoradiotherapy in locally advanced hepatocellular carcinoma patients.docx

Background and purposeHepatocellular carcinoma (HCC) is the third leading cause of cancer-related mortality, and intrahepatic progression after following treatment is common. Accurate tumor evaluation is essential for treatment decisions but remains challenging due to tumor heterogeneity, the background of cirrhotic liver, and treatment-related artifacts. This study investigated the feasibility of a deep learning–based auto-segmentation approach for response evaluation in locally advanced HCC treated with concurrent chemoradiotherapy (CCRT).
MethodsWe retrospectively analyzed 83 treatment-naïve patients with locally advanced HCC who underwent definitive CCRT between 2016 and 2021. Tumor contours were manually delineated on pre-treatment (CTpre) and first post-treatment CT (CTpost). A fully convolutional DenseNet (FCD) and an intentional deep overfit learning (IDOL) framework were trained and validated. Performance was assessed using the Dice similarity coefficient (DSC), and RECIST-based diameters were compared between manual and predicted contours.
ResultsIn the full cohort, the FCD model achieved mean DSCs of 0.53 for CTpre and 0.33 for CTpost, while the IDOL model improved CTpost DSCs to 0.49. In the RECIST cohort (n = 63), mean DSCs were 0.61 for CTpre and 0.53 for CTpost using FCD, versus 0.63 for IDOL. For the RECIST cohort (n = 14 validation cases), predicted diameters differed by a mean of 9.2 mm from manual values (p = 0.032), showing a tendency toward overestimation in peritumoral inflammatory areas. However, RECIST-based response showed high concordance in 13 of 14 cases.
ConclusionsThe patient-specific IDOL framework improved auto-segmentation accuracy compared with conventional models and provided reliable data for RECIST-based response assessment. Despite limitations and lack of external validation, this study demonstrates the preliminary feasibility of auto-segmentation to support response evaluation in treated HCC.