Oct, 2023

自主训练的掩蔽关注引导的掩蔽图像建模与噪音约束教师 (SMART) 用于医学图像分析

TL;DRHierarchical shifted window transformers (Swin) were architecturally enhanced with semantic class attention for self-supervised attention guided co-distillation with masked image modeling (MIM), resulting in SMART. SMART, pretrained with 10,412 unlabeled 3D computed tomography (CTs), demonstrated high performance in multiple downstream tasks involving lung cancer (LC) analysis, including predicting immunotherapy response, LC recurrence, LC segmentation, and unsupervised clustering of organs in the chest and abdomen, without finetuning.