Integrating habitat-based radiomics, 3D deep learning, and clinical variables for detection of early-stage esophageal squamous cell carcinoma on noncontrast chest CT: A multicenter study.

RATIONALE AND OBJECTIVES: To develop and validate a Combined model integrating habitat-based sub-regional radiomics, 3D deep learning (DL) features, and clinical variables for the detection of early-stage esophageal squamous cell carcinoma (ESCC) on noncontrast chest CT.

MATERIALS AND METHODS: In this retrospective multicenter study, the segmented esophageal region of interest was partitioned into habitat sub-regions via unsupervised clustering, and radiomic features extracted from these sub-regions were fused with deep features derived from a 3D ResNet-101 to construct a deep learning-radiomics fusion model (DLR). Selected clinical variables were further integrated with the DLR signature to form the Combined model. Performance was evaluated by the area under the curve (AUC), calibration curves, and decision curve analysis (DCA). Model interpretability was assessed using SHAP analysis, and an exploratory reader-comparison study evaluated the effect of AI assistance on radiologist performance.

RESULTS: A total of 469 subjects were included. The DLR model outperformed both the habitat-based radiomic model and the standalone DL model across cohorts. The Combined model achieved the highest AUC of 0.895 (internal validation) and 0.847 (external test), with favorable calibration and DCA. AI assistance improved the sensitivity of all four radiologists for early-stage ESCC and increased inter-reader agreement.

CONCLUSION: In this retrospective case-control study, the Combined model demonstrated promising discrimination and calibration for early-stage ESCC, with preliminary cross-center performance. These findings support prospective evaluation of the model as an opportunistic second-reader and triage tool for already-acquired noncontrast chest CT examinations.