[3Xin2] Poster session 1

[3Xin2-75] Multi-task Representation Learning for Medical Image Analysis

〇Eichi Takaya¹, Ryusei Inamori² (1.Tohoku University Hospital, 2.Tohoku University Graduate School of Medicine)

Keywords:Deep Learning, Representation Learning, Medical Image

In the diagnosis and treatment planning of diseases based on medical imaging, the analysis of fine features in images is crucial, requiring a deep understanding of the photographed region and anatomical structures. This necessitates a comprehensive knowledge of various imaging modalities and conditions. In the context of medical AI, there is a demand for the development of new learning methods that more effectively incorporate these elements in medical image analysis. This study proposes a multitask representation learning approach that combines hierarchical contrastive learning aimed at acquiring unique knowledge in medical imaging and a masked auto-encoder to deepen the understanding of anatomical structures. By solving two types of pre-training tasks simultaneously, it is expected to acquire representations that consider medical imaging-specific knowledge and anatomical structures, and to improve accuracy in downstream tasks. For evaluation, pre-training is conducted using RadImageNet, a large-scale medical image dataset that includes three modalities (CT, MR, US) and a variety of anatomical structures. The effectiveness of the proposed method is demonstrated by measuring performance in the thyroid nodule classification task using a relatively small dataset.