KnowMIM: a Self-supervised Pre-training Framework Based on Knowledge-Guided Masked Image Modeling for Retinal Vessel Segmentation

Mainstream segmentation algorithms currently rely on supervised learning and thus require large pixel-labelled datasets for training. However, manually labelling regions of interest in medical images is both time-consuming and expertise-demanding, compressing the scale of the dataset and thus limiting the accuracy of medical image segmentation. Self-supervised learning is often preferred in cases where annotation dependency needs to be alleviated or budget is limited. However, existing methods have ignored the properties of medical images and lack adaptive masking methods, resulting in poor generalisation. This paper proposes a self-supervised pre-training framework for retinal vessel segmentation based on a priori knowledge. The proposed framework is called KnowMIM and works in two phases guided with knowledge: (1) Adaptive masks generation. KnowMIM utilises an edge detection algorithm to extract the location of vessel contours as a priori information, and then generates an adaptive mask for each retinal image, which is masked for data augmentation, and (2) Masked image modeling. KnowMIM carries out masked image modeling via a U-Net architecture and performs reconstruction through self-supervised learning to pre-train the encoder and decoder. Extensive experiments have been conducted on public retinal datasets for vessels segmentation. Results demonstrate that KnowMIM outperforms state-of-the-art pre-training counterparts. Additionally, KnowMIM effectively utilises unlabelled data and exhibits generalisation on external datasets.