AI model development
Our AI model was developed based on the algorithm of YOLO (you only look once, V3) , a unified, real-time, efficient object detection algorithm, which was recently proposed in deep learning computer vision field24-26. Object detection algorithms were designed not only to recognize what objects are present but also to localize where they are, no matter how many objects are there. Thus, object detection is more complex and challenging compared with classification algorithms. It was initially used in face recognition in security field and self-driving. In the ultrasound imaging field, there might be unknown number of structures and lesions within one image that need to be recognized and precisely located. Also, we chose YOLO for its efficiency considering dynamic data analysis may be needed. We added a logic output network to YOLO in our current AI model, which would eliminate redundant labels on the same structure by comparing label scores. For example, for the same image, normal and abnormal labels could not simultaneously exit on the same side of the lateral ventricles. As a result, the model had only one input and two outputs. The input of the model was the ultrasound image of fetal brain. The first output was a bounding box with labels and scores (numbers range from 0 to 1). The second output was the final result which consisted of remaining bounding boxes with labels after label elimination in the logic output, as shown in Figure 2 and Figure 3. Note that, due to the logic output network, lesions detected by AI were not made only based on label scores which were continuous number from 0 to 1 but also on the higher score. Therefore, when we drew ROC, the data were treated as binary data (yes/no) like human making diagnosis, rather than continuous variable data.