A dynamic few-shot learning framework for medical image stream mining based on self-training

Abstract Few-shot semantic segmentation (FSS) has been widely used in the field of information medicine and intelligent diagnosis. Due to high cost medical data collection privacy protection patients, labeled images are difficult obtain. Compared with other dataset which can be automatically generated a large scale, image tend continually generated. Most existing FSS techniques require abundant annotated classes for pre-training cannot deal its dynamic nature stream. To this issue, we propose few-shot learning framework segmentation, fully utilize features newly-collected/generated We introduce new pseudo-label generation strategy continuously generating pseudo-labels avoiding model collapse during self-training. Furthermore, an efficient consistency regularization is proposed limited data. The iteratively trained on three tasks: abdominal organ CT MRI, cardiac MRI. Experiments results demonstrate significant performance gain stream mining compared baseline method.