ESTRO 2024 - Abstract Book

S4411

Physics - Machine learning models and clinical applications

ESTRO 2024

36

Digital Poster

deep learning method for classification of congenital and secondary biliary dilatation

Jinghua Yue 1 , Nan Jiang 2 , Fugen Zhou 1 , Siyuan Wang 2 , Bin Guo 1 , Shaoqing Yu 2 , Benqi Zhao 3 , Jianping Zeng 2 , Bo Liu 1 , Shuo Jin 2 1 Beihang University, Image Processing Center, Beijing, China. 2 Tsinghua University, Beijing Tsinghua Changgung Hospital, School of Clinical Medicine, Beijing, China. 3 Tsinghua University, Beijing Tsinghua Changgung Hospital, Department o f Radiology, Beijing, China

Purpose/Objective:

Congenital biliary dilatation (CBD) primarily occurs in children, and its incidence is increasing in adults [1,2] . The complication rates of CBD are approximately 20%–90% [3,4] . The most fatal complication is biliary carcinogenesis. Some studies even suggest that CBD can be considered a premalignant condition [5] . Therefore, the treatment principle for CBD is timely surgical resection after early diagnosis [6,7] . In contrast to CBD, which is considered a primary biliary dilatation, secondary biliary dilatation (SBD) can be caused by the obstruction of the bile duct, owing to pancreatic cancer, terminal cholangiocarcinoma, hilar cholangiocarcinoma (HCCA), and cholangiolithiasis. The clinical classification of CBD and SBD relies on the expertise and experience of radiologists and hepatobiliary surgeons; however, a high rate of misdiagnosis occurs. Thus, the objective and accurate diagnosis of CBD is particularly important, because their treatment and prognosis are completely different. To the best of our knowledge, no previous studies have reported the use of the deep learning method to distinguish CBD from SBD. Thus, this study aimed to develop an objective and accurate deep learning approach for the classification of CBD and SBD. A total of 310 dynamic enhanced computed tomography (CT) images were retrospectively collected from 285 patients diagnosed with either CBD or SBD. Five biliary dilatation categories were included in the collected dataset: category A, CBD prior to biliary drainage; category B, CBD after biliary drainage; category C, SBD caused by tumor (including pancreatic cancer and terminal cholangiocarcinoma); category D, SBD caused by extrahepatic and intrahepatic cholangiolithiasis; and category E, SBD caused by HCCA. The data preprocessing phase included resampling both the CT images and annotations, extracting regions of interest (ROIs), applying normalization, and performing data enhancement. A three-dimensional convolutional neural network (3D CNN) model was proposed that used the intensity and morphology information (annotation mask) of the biliary tract as input, owing to the complex structure and low intensity contrast of the biliary tract. The model consists of ten convolutional layers, four maximum pooling layers, and a global average pooling layer. The global average pooling layer was used to replace the traditional fully connected layer, which considerably reduces the network parameters and allows input of different sizes. The model results were compared with those obtained from clinicians. Two attending hepatobiliary surgeons were selected over radiologists to participate in the reader study validation, in order to more accurately emulate the real clinical decision making process. Several variants of the proposed method and training data strategies were also investigated to analyze the method design and the impact of training data categories. A total of 150 CT images of different categories were used as test datasets for model evaluation. Metrics including accuracy, specificity, sensitivity, and the area under the curve (AUC) were used for model evaluation. Material/Methods: