Combining self-attention and convolution to understand OCT images

In the application of AI-enabled medical image analysis, when correctly clinically annotated data for training is difficult to acquire, we propose Domain Adaptation (DA) techniques as solution. This paper aims to provide an analysis of deep neural network (DNN)-based DA techniques for medical image segmentation towards clinical diagnostics given different levels of available annotation in target domain. DA techniques using supervised, unsupervised, and a novel semi-supervised approach were compared to minimize the discrepancy between datasets with large domain shifts for medical image segmentation application. The domain shifts were introduced not only from the different instruments and datasets but also different body organs and imaging modalities. The DA techniques were evaluated on DNN models which were used to segment blood vessels from Colour Fundus Photography retina images as the target domain, given X-ray Coronary Angiography images as the source domain or training data. Fine-tuning was used for supervised DA and Mean-teacher was used for unsupervised DA. A novel multi-source Mean-teacher (MSMT) was developed as semi-supervised DA. Results show that fine-tuning using a few number of target domain data, or the so-called few-shots fine-tuning technique, provides a simple and efficient solutions.<br/><br/>