Multi-instance Learning as Downstream Task of Self-Supervised Learningbased Pre-trained Model

Multi-instance learning (MIL) is a type of weakly supervised learning. In MIL, a set of instances is defined asa bag, and only bag labels are known. In recent years, deep learning-based MIL has been applied in medical imaging,especially in histopathological images. However, when classifying the marker in the brain hematoma CT imageswhich we deal with, simply applying MIL leads to spurious correlations. In addition, many studies applying MIL tohistopathological images handle thousands of instances per bag and show high performance, while our task strugglesto learn with a few hundred instances per bag and has low performance. In this paper, we address these problems. Wepropose using a pre-trained model with self-supervised learning for the multi-instance learner as a downstream task.We apply self-supervised learning to the instance feature extractor in MIL to learn effective feature representations ofpatch images. The experimental results demonstrate that our method achieves improvements of 6.5% in accuracy and11.2% in the F1-score without falling into the spurious correlation problem.