Machine-learning radiomics to predict early recurrence in perihilar cholangiocarcinoma after curative resection.

BACKGROUND AND AIMS: Up to 40%-65% of patients with perihilar cholangiocarcinoma (PHC) rapidly progress to early recurrence (ER) even after curative resection. Quantification of ER risk is difficult and a reliable prognostic prediction tool is absent. We developed and validated a multilevel model, integrating clinicopathology, molecular pathology and radiology, especially radiomics coupled with machine-learning algorithms, to predict the ER of patients after curative resection in PHC.

METHODS: In total, 274 patients who underwent contrast-enhanced CT (CECT) and curative resection at 2 institutions were retrospectively identified and randomly divided into training (n = 167), internal validation (n = 70) and external validation (n = 37) sets. A machine-learning analysis of 18,120 radiomic features based on multiphase CECT and 48 clinico-radiologic characteristics was performed for the multilevel model.

RESULTS: Comprehensively, 7 independent factors (tumour differentiation, lymph node metastasis, pre-operative CA19-9 level, enhancement pattern, A-Shrink score, V-Shrink score and P-Shrink score) were built to the multilevel model and quantified the risk of ER. We benchmarked the gain in discrimination with the area under the curve (AUC) of 0.883, superior to the rival clinical and radiomic models (AUCs 0.792-0.805). The accuracy (ACC) of the multilevel model was 0.826, which was significantly higher than those of the conventional staging systems (AJCC 8th (0.641), MSKCC (0.617) and Gazzaniga (0.581)).

CONCLUSION: The radiomics-based multilevel model demonstrated superior performance to rival models and conventional staging systems, and could serve as a visual prognostic tool to plan surveillance of ER and guide post-operative individualized management in PHC.