Self-Supervised Learning for Accurate Liver View Classification in Ultrasound Images With Minimal Labeled Data
Conventional B-mode "grey scale" medical ultrasound and shear wave elastography (SWE) are widely used for chronic liver disease diagnosis and risk stratification. Liver disease is very common and is clinically and socially important. As a result, multiple medical device manufacturers have proposed or developed AI systems for ultrasound image analysis. However, many abdominal ultrasound images do not include views of the liver, necessitating manual data curation for model development. To optimize the efficiency of real-time processing, a pre-processing liver view detection step is necessary before feeding the image to the AI system. Deep learning techniques have shown great promise for image classification, yet labeling large datasets for training classification models is time-consuming and expensive. In this paper, we present a self-supervised learning method for image classification that utilizes a large set of unlabeled abdominal ultrasound images to learn image representations. These representations are then applied on the downstream task of liver view classification, resulting in efficient classification and alleviation of the labeling burden. In comparison to two state-of- the-art (SOTA) models, ResNet-18 and MLP-Mixer, when trained for 100 epochs the proposed SimCLR+LR approach demonstrated outstanding performance when only labeling "one" image per class, achieving an accuracy similar to MLP-Mixer (86%) and outperforming the performance of ResNet-18 (70.2%), when trained on 854 (with liver: 495, without liver: 359) B-mode images. When trained on the whole dataset for 1000 epochs, SimCLR+LR and ResNet- 18 achieved an accuracy of 98.7% and 79.3%, respectively. These findings highlight the potential of the SimCLR+LR approach as a superior alternative to traditional supervised learning methods for liver view classification. Our proposed method has the ability to reduce both the time and cost associated with data labeling, as it eliminates the need for human labor.