Self- and Semi-supervised Learning for Gastroscopic Lesion Detection

Gastroscopic Lesion Detection (GLD) plays a key role in computer-assisted diagnostic procedures. However, this task is not well studied in the literature due to the lack of labeled data and the applicable methods. Generic detectors perform below expectations on GLD tasks for 2 reasons: 1) The scale of labeled data of GLD datasets is far smaller than that of natural-image object detection datasets. 2) Gastroscopic images exhibit distinct differences from natural images, which are usually of high similarity in global but high diversity in local. Such characteristic of gastroscopic images also degrades the performance of using generic self-supervised or semi-supervised methods to solve the labeled data shortage problem using massive unlabeled data. In this paper, we propose Self- and Semi-Supervised Learning (SSL) for GLD tailored for using massive unlabeled gastroscopic images to enhance GLD tasks performance, which consists of a Hybrid Self-Supervised Learning (HSL) method for backbone pre-training and a Prototype-based Pseudo-label Generation (PPG) method for semi-supervised detector training. The HSL combines patch reconstruction with dense contrastive learning to boost their advantages in feature learning from massive unlabeled data. The PGG generates pseudo-labels for unlabeled data based on similarity to the prototype feature vector to discover potential lesions and avoid introducing much noise. Moreover, we contribute the first Large-scale GLD Datasets (LGLDD), which contains 10,083 gastroscopic images with 12,292 well-annotated boxes for four-category lesions. Experiments on LGLDD demonstrate that SSL can bring significant improvement.