Evaluate the diagnostic performance of a deep learning model in assisting primary care ultrasound doctors in differentiating benign and malignant breast nodules.

This study retrospectively collected female patients who were diagnosed with BI-RADS category 3 and 4 breast nodules during routine ultrasound examinations at Shanghai Sixth People's Hospital from September 2022 to September 2024. The clinical data and ultrasound images of these patients were collected.. Five deep learning models (GoogLeNet, VGG11, AlexNet, ResNet-34, ViT) were constructed, and the optimal model was selected based on metrics such as area under curve (AUC). Visual heatmaps of the ultrasound images were generated using the optimal model. Community ultrasound physicians independently reviewed the images without knowledge of the pathological results. Two weeks later, the physicians re-reviewed the images with the assistance of the generated heatmaps. Diagnostic performance was compared among three approaches: deep learning model alone, physician independent review, and deep learning-assisted physician review, using accuracy, sensitivity, specificity, and precision as metrics.

A total of 430 patients with breast nodules were included and divided into a training set (n=301) and a validation set (n=129). Among the different learning models constructed based on breast ultrasound images, the ResNet-34 model performed best in predicting benign and malignant breast nodules, with AUC, accuracy, and sensitivity superior to other models in both the training and validation sets. Therefore, the ResNet-34 model was selected as the optimal model. Among the three approaches, the diagnostic efficacy of independent review by primary care physicians was the lowest. The deep learning model slightly outperformed primary care physicians in terms of accuracy, sensitivity, and precision, while the diagnostic efficacy of the deep learning model assisting ultrasound physicians in reading was the strongest, with all evaluation indicators significantly improved. In terms of the improvement rate, the sensitivity and F1 value showed the most significant increase, rising by 34.3% and 29.9% respectively.

Deep learning models based on ultrasound images can effectively distinguish the benign and malignant nature of breast nodules. Combined with the generated visual heatmaps, they can significantly enhance the diagnostic efficacy of assisting primary ultrasound physicians in distinguishing between benign and malignant breast nodules, providing valuable guidance for selecting treatment plans for patients with breast nodules.