Using deep learning models in magnetic resonance cholangiopancreatography images to diagnose common bile duct stones

Abstract

Backgrounds and aims

Magnetic resonance cholangiopancreatography (MRCP) plays a significant role in diagnosing common bile duct stones (CBDS). Currently, there are no studies to detect CBDS by using the deep learning (DL) model in MRCP. This study aimed to use the DL model You Only Look Once version 5 (YOLOv5) to diagnose CBDS in MRCP images and verify its validity compared to the accuracy of radiologists.

Methods

By collecting the thick-slab MRCP images of patients diagnosed with CBDS, 4 submodels of YOLOv5 were used to train and validate the performance. Precision, recall rate, and mean average precision (mAP) were used to evaluate model performance. Analyze possible reasons that may affect detection accuracy by validating MRCP images in 63 CBDS patients and comparing them with radiologist detection accuracy. Calculate the correctness of YOLOv5 for detecting one CBDS and multiple CBDS separately.

Results

The precision of YOLOv5l (0.970) was higher than that of YOLOv5x (0.909), YOLOv5m (0.874), and YOLOv5s (0.939). The mAP did not differ significantly between the 4 submodels, with the following results: YOLOv5l (0.942), YOLOv5x (0.947), YOLO5s (0.927), and YOLOv5m (0.946). However, in terms of training time, YOLOv5s was the fastest (4.8 h), detecting CBDS in only 7.2 milliseconds per image. In 63 patients the YOLOv5l model detected CBDS with an accuracy of 90.5% compared to 92.1% for radiologists, analyzing the difference between the positive group successfully identified and the unidentified negative group not. The incorporated variables include common bile duct diameter > 1 cm (p = .560), combined gallbladder stones (p = .706), maximum stone diameter (p = .057), combined cholangitis (p = .846), and combined pancreatitis (p = .656), and the number of CBDS (p = .415). When only one CBDS was present, the accuracy rate reached 94%. When multiple CBDSs were present, the recognition rate dropped to 70%.

Conclusion

YOLOv5l is the model with the best results and is almost as accurate as the radiologist’s detection of CBDS and is also capable of detecting the number of CBDS. Although the accuracy of the test gradually decreases as the number of stones increases, it can still be useful for the clinician’s initial diagnosis.

Keywords:

• Deep learning
• magnetic resonance cholangiopancreatography
• common bile duct
• stone

Ethical approval

The Institutional Review Board of the First Affiliated Hospital of Chengdu Medical College approved our study. (2023CYFYIRB-BA-JuI04).

Author contributions

B.L, Z.L, and J.H contributed to the conception of the study, B.L and J.H analyzed the data and wrote the manuscript, K.Z, S.K, W.C, N.F, and Z.Y collected the data and provided constructive discussions.

Disclosure statement

Prof. Jingcheng Hao, Mr. Bo Luo, Mr. Zhiyuan Li, Mr. Ke Zhan, Mr. Sikai Wu, Dr. Weiwei Chen, Dr. Ning Fu, and Prof. Zhiming Yang have no conflicts of interest or financial ties to disclose.

Additional information

Funding

None