Abstract No. 499 Novel needle guidance technology in the angiography room: from CBCT to stereotaxic reconstruction

Abstract

However, the procedure is invasive and carries the risk of infection and bleeding. In this study, we used deep learning to differentiate benign from malignant focal solid liver lesions based on their appearance on routine abdominal ultrasound. Materials: A ResNet50-based neural network was fine-tuned using pre-trained weights on ImageNet. Patients who had abdominal ultrasound from 2014-2018 with liver code C2-C5 on Code Abdomen, an institutional lesion categorization system for abdominal organs on cross-sectional imaging modeled after BI-RADS, were included (Table 1). Non-cystic liver lesions with definite diagnosis by histopathology or MRI were included. Our dataset was split into training and validation sets by patient and augmented in real-time with zoom, skew, and flip transformations during training. Subgroup analysis of category 3 and 4 lesions (uncertain diagnosis set) was performed. Accuracy of the final model was compared with expert interpretation. Results: Among the 410 patients who met the inclusion criteria, there were 521 images of individual lesions, of which 170 were malignant and 351 were benign. Our training set had 419 lesions and our validation set had 102 lesions. Our model achieved a validation accuracy of 83.3% compared to expert accuracy of 77.5% and 60.8%, respectively. The model had an ROC AUC of 0.83, sensitivity of 0.82, and specificity of 0.84. For the uncertain diagnosis set, our model achieved a validation accuracy of 85.1% compared to expert accuracy of 74.5% and 61.7%, respectively. The model had an ROC AUC of 0.78, sensitivity of 0.91, and specificity of 0.67. Conclusions: Deep learning distinguished benign from malignant ultrasound-captured solid liver lesions with a high accuracy compared to expert radiologists and can better triage patients for biopsy, particularly those at high risk or with multiple lesions.