Deep learning-based classification of gallbladder lesions in patients with non-diagnostic (GB-RADS 0) ultrasound.

Aim of the study: Non-diagnostic ultrasound (US) of the gallbladder may be due to various factors. We aimed to evaluate the diagnostic performance of deep learning-based classification of gallbladder lesions on US images in patients with non-diagnostic US secondary to gallbladder factors. Material and methods: Consecutive patients with non-diagnostic US due to calculi within the gallbladder lumen, obscuring the detailed evaluation, were identified by a research fellow from a prospective database of patients with gallbladder lesions. The US reports and images were evaluated by a radiologist blinded to the final diagnosis. Patients who had the final pathological diagnosis based on fine-needle aspiration cytology, percutaneous or endoscopic biopsy, or surgical histopathology were included. Convolution neural networks (ResNet50, GBCNet), transformer models (vision transformer [ViT], RadFormer), and a hybrid model (MedViT) were trained on a public gallbladder dataset (GBCU dataset). The performance of these models for classifying gallbladder lesions into benign and malignant was tested on non-diagnostic (GB-RADS 0) US images. Results: Training and validation cohorts (GBCU dataset) comprised 1004 and 251 images, respectively. The testing data (26 patients, mean age [SD]: 57.5 ±8.07 years, 17 female) comprised 304 images. The best performance for detection of GBC was achieved with GBCNet (sensitivity 51.1%, specificity 83.3%, area under the curve [AUC] 0.709) and MedViT (sensitivity 92.8%, specificity 50%, AUC 0.714). MedViT had the best accuracy (73.1%) for detecting benign gallbladder lesions. Conclusions: These results suggest that deep learning models can potentially stratify patients with non-diagnostic US. However, further improvement in the performance is needed to render this approach relevant in clinical practice.