Radiomics-based Deep Learning Approach to Predict Recurrence in Cholangiocarcinoma Patients after Curative Resection: A Pilot Study

Prin Twinprai, MD^1,2, Piyakarn Watcharenwong, MD^2,3, Warunporn Siriboonpipattana, MD^2,3, Attapol Titapun, MD,PhD^2,4, Vasin Thanasukarn, MD^2,4, Nattaphon Twinprai, MD⁵, Prinya Chindaprasirt, PhD⁶, Puripong Suthisopapan, PhD⁶, Jarin Chindaprasirt, MD^2,3

Affiliation : ¹ Department of Radiology, Faculty of Medicine, Khon Kaen University, Khon Kaen, Thailand,² Cholangiocacinoma Research Institute, Khon Kaen, Thailand, ³ Division of Medical Oncology, Department of Medicine, Faculty of Medicine, Khon Kaen University, Khon Kaen, Thailand, ⁴ Division of Hepatobiliary Surgery, Department of Surgery, Faculty of Medicine, Khon Kaen University, Khon Kaen, Thailand,⁵ Department of Orthopedics, Faculty of Medicine, Khon Kaen University, Khon Kaen, Thailand,⁶ Faculty of Engineering, Khon Kaen University, Khon Kaen, Thailand

Objective: The purpose of the present study was to develop a preoperative CT-radiomics model for predicting recurrence for patients with cholangiocarcinoma (CCA) after curative resection.
Materials and Methods: Preoperative contrast-enhanced CT scans from a randomized phase 3 CCA trial were included (n=56). Segmentation was performed using 3DSlicer program and radiomics-based features were extracted using PyRadiomics. Feature selection was performed by using XGBoost and SHAP values to identify the most important features which consequently were used to train a classification neural network in MATLAB for predicting recurrence.
Results: With a median follow-up time of 55 months, 29 patients had disease recurrence. The 1-yr, 2-yr, and 5-yr recurrence-free survival time rates were 76.8%, 60.7%, and 41.4%, respectively. The ten most important features predicting recurrence were selected and the radiomics-based model was trained for 1,000 epochs and augmented three times, achieving a sensitivity and specificity of 1.0 and accuracy of 1.0.
Conclusion: Radiomics-based deep learning could serve as a valuable tool for predicting the recurrence of cholangiocarcinoma patients in a preoperative setting.

Keywords : Cholangiocarcinoma; Radiomics; Recurrence; Deep learning; XGBoost

All Articles Download

Information

Related Links