Deep learning for differentiation of benign and malignant solid liver lesions on ultrasonography

• Published in: Abdominal imaging 2021-02, Vol.46 (2), p.534-543
• Main Authors: Xi, Ianto Lin, Wu, Jing, Guan, Jing, Zhang, Paul J., Horii, Steven C., Soulen, Michael C., Zhang, Zishu, Bai, Harrison X.
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Benign; Biopsy; Datasets; Deep learning; Diagnosis; Differentiation; Gastroenterology; Hepatobiliary; Hepatology; Histopathology; Imaging; Learning algorithms; Lesions; Liver; Liver cancer; Machine learning; Magnetic resonance imaging; Medical diagnosis; Medical imaging; Medicine; Medicine & Public Health; Model accuracy; Model testing; Neural networks; Radiology; Training; Ultrasonic imaging; Ultrasonic testing; Ultrasound
• ISSN: 2366-004X; 2366-0058
• DOI: 10.1007/s00261-020-02564-w

Purpose The ability to reliably distinguish benign from malignant solid liver lesions on ultrasonography can increase access, decrease costs, and help to better triage patients for biopsy. In this study, we used deep learning to differentiate benign from malignant focal solid liver lesions based on their ultrasound appearance. Methods Among the 596 patients who met the inclusion criteria, there were 911 images of individual liver lesions, of which 535 were malignant and 376 were benign. Our training set contained 660 lesions augmented dynamically during training for a total of 330,000 images; our test set contained 79 images. A neural network with ResNet50 architecture was fine-tuned using pre-trained weights on ImageNet. Non-cystic liver lesions with definite diagnosis by histopathology or MRI were included. Accuracy of the final model was compared with expert interpretation. Two separate datasets were used in training and evaluation, one with all lesions and one with lesions deemed to be of uncertain diagnosis based on the Code Abdomen rating system. Results Our model trained on the complete set of all lesions achieved a test accuracy of 0.84 (95% CI 0.74–0.90) compared to expert 1 with a test accuracy of 0.80 (95% CI 0.70–0.87) and expert 2 with a test accuracy of 0.73 (95% CI 0.63–0.82). Our model trained on the uncertain set of lesions achieved a test accuracy of 0.79 (95% CI 0.69–0.87) compared to expert 1 with a test accuracy of 0.70 (95% CI 0.59–0.78) and expert 2 with a test accuracy of 0.66 (95% CI 0.55–0.75). On the uncertain dataset, compared to all experts averaged, the model had higher test accuracy (0.79 vs. 0.68, p  = 0.025). Conclusion Deep learning algorithms proposed in the current study improve differentiation of benign from malignant ultrasound-captured solid liver lesions and perform comparably to expert radiologists. Deep learning tools can potentially be used to improve the accuracy and efficiency of clinical workflows.