Background: The differential diagnosis between gallbladder cancer (GBC) and xanthogranulomatous cholecystitis (XGC) remains quite challenging, and can possibly lead to improper surgery. This study aimed to distinguish between XGC and GBC by combining computed tomography (CT) images and deep learning (DL) to maximize the therapeutic success of surgery.
Methods: We collected a dataset, including preoperative CT images, from 28 cases of GBC and 21 XGC patients undergoing surgery at our facility. It was subdivided into training and validation (n = 40), and test (n = 9) datasets. We built a CT patch-based discriminating model using a residual convolutional neural network and employed 5-fold cross-validation. The discriminating performance of the model was analyzed in the test dataset.
Results: Of the 40 patients in the training dataset, GBC and XGC were observed in 21 (52.5%), and 19 (47.5%) patients, respectively. A total of 61 126 patches were extracted from the 40 patients. In the validation dataset, the average sensitivity, specificity, and accuracy were 98.8%, 98.0%, and 98.5%, respectively. Furthermore, the area under the receiver operating characteristic curve (AUC) was 0.9985. In the test dataset, which included 11 738 patches, the discriminating accuracy for GBC patients after neoadjuvant chemotherapy (NAC) (n = 3) was insufficient (61.8%). However, the discriminating model demonstrated high accuracy (98.2%) and AUC (0.9893) for cases other than those receiving NAC.
Conclusion: Our CT-based DL model exhibited high discriminating performance in patients with GBC and XGC. Our study proposes a novel concept for selecting the appropriate procedure and avoiding unnecessary invasive measures.
Keywords: deep learning; gallbladder cancer; neural network; precision medicine; xanthogranulomatous cholecystitis.
© 2022 The Authors. Annals of Gastroenterological Surgery published by John Wiley & Sons Australia, Ltd on behalf of The Japanese Society of Gastroenterology.