Applications of Automated Machine Learning Diagnosis in Medical Ultrasound

While medical ultrasound has been a well poised imaging technique for the early detection and diagnosis of breast cancer, thyroid cancer, and fatty liver diseases, it has suboptimal operator and reader dependence. The use of machine learning in ultrasound-based detection and diagnosis has shown an ability to increase diagnostic accuracy, decrease time and resources requirements for diagnosis, and reduce reader variability. Hence, this project investigated the use of Google AutoML Vision (Google, Mountain View, CA), a commercially available machine learning service, to characterize indeterminate breast lesions, thyroid nodules, and fatty liver disease on B-mode ultrasound. The performances of two sub-models from AutoML Vision, the Image Classification Model and the Object Detection Model were evaluated, while also investigating training strategies to enhance model performance. For breast cancer identification and classification, the best model from the Image Classification Model had an AUC of 0.91 and 76.9% accuracy of prediction testing, while training under unbalanced training condition (147 benign vs 117 malignant). The Object Detection Model on breast lesion identification had an AUC of 0.74 with only 67% accurate from prediction testing but showed the location of the lesions and provided a higher percent confidence in correctly classified lesions (98%) than the Image Classification Model (83%). The thyroid nodules Object Detection Model had an AUC of 0.89 during model validation and subsequently classified 34 out of 51 images correctly (66.7% prediction accuracy) with 100% lesion detection. For fatty liver classification, data accrual is ongoing, but to date the model has an AUC of 0.88 during model validation and classified 29 out of 40 images correctly (72.5% prediction accuracy) when used for prospective prediction. In conclusion, the two models appear to be useful tools to assist radiologists in classifying and identifying suspicious areas on breast, thyroid, and liver ultrasound.