Endoscopic ultrasound diagnosis system based on deep learning in images capture and segmentation training of solid pancreatic masses

• Published in: Medical physics (Lancaster) Vol. 50; no. 7; pp. 4197 - 4205
• Main Authors: Tang, Anliu, Gong, Pan, Fang, Ning, Ye, Mingmei, Hu, Shan, Liu, Jinzhu, Wang, Wujun, Gao, Kui, Wang, Xiaoyan, Tian, Li
• Format: Journal Article
• Language: English
• Published: United States 01.07.2023
• Subjects: contrast‐enhanced harmonic endoscopic ultrasound; Deep Learning; Endosonography - methods; Humans; images capture and segmentation; Pancreas - diagnostic imaging; Pancreatic Neoplasms - diagnostic imaging; Retrospective Studies; solid pancreatic masses; training; deep learning; contrast-enhanced harmonic endoscopic ultrasound; training; images capture and segmentation; solid pancreatic masses
• DOI: 10.1002/mp.16390
• ISSN: 0094-2405; 2473-4209

Background Early detection of solid pancreatic masses through contrast‐enhanced harmonic endoscopic ultrasound (CH‐EUS) is important. But CH‐EUS is difficult to learn. Purpose To design a deep learning‐based CH‐EUS diagnosis system (CH‐EUS MASTER) for real‐time capture and segmentation of solid pancreatic masses and to verify its value in the training of pancreatic mass identification under endoscopic ultrasound (EUS). Methods We designed a real‐time capture and segmentation model for solid pancreatic masses and then collected 4530 EUS images of pancreatic masses retrospectively, used for training and testing of this model at a ratio of 8:2. The model is loaded into the EUS host computer to establish the CH‐EUS MASTER system. A crossover trial was then conducted to evaluate the model's value in EUS trainee training by successfully conducting two groups of EUS trainees in model learning and trainer‐guided training. The intersection over union (IoU) and the time to find the lesion were used to evaluate the model performance metrics, and the Mann–Whitney test was used to compare the IoU and the time to find the lesion in different groups of subjects. Paired t‐test was used to compare the effects before and after training. When α ≤ 0.05, it is considered to have a significant statistical difference. Results The model test showed that the model successfully captured and segmented the pancreatic solid mass region in 906 EUS images. The real‐time capture and segmentation model had a Dice coefficient of 0.763, a recall rate of 0.941, a precision rate of 0.642, and an accuracy of 0.842 (when the threshold is set to 0.5), and the median IoU of all cases was 0.731. For the AI training effect, the average IoU of eight trainees improved from 0.80 to 0.87 (95% CI, 0.032–0.096; p = 0.002). The average time for identifying lesions in the pancreatic body and tail improved from 22.75 to 17.98 s (95% CI, 3.664–5.886; p < 0.01). The average time for identifying lesions in the pancreatic head and uncinate process improved from 34.21 to 25.92 s (95% CI, 7.661–8.913; p < 0.01). Conclusion CH‐EUS MASTER can provide an effect equivalent to trainer guidance in training pancreatic solid mass identification and segmentation under EUS.