어탠션 메커니즘을 이용한 강인한 표현 학습 기반 추적 흉부엑스선 변화 평가 인공지능 시스템 개발에 대한 연구

Abstract

Diagnosing meaningful changes in follow-up CXRs which is one of the main tasks of radiologists in routine clinical practice is challenging because radiologists must distinguish between pathological changes and natural or benign changes and inevitably it requires a large amount of datasets and high quality annotation. However, it is very difficult to acquire those kinds of high-quality annotation, in actual clinical setting. In this paper, a multi-task Siamese convolutional vision transformer (MuSiC-ViT) with an anatomy matching module (AMM) is proposed to mimic the radiologist's cognitive process in classifying CXR pairs (baseline change/no change) and follow-up radiographs to overcome these issues. MuSiC-ViT uses the CNNs and visual transformers (CMTs) model, which combines the CNN and transformer architecture and consists of three main components: a Siamese architecture, an anatomy matching module, and multi-task learning. Since the input was a pair of CXRs, a Siamese network was chosen for the encoder network. The AMM is an attentional module that focuses on related regions in corresponding CXR pairs. To mimic the cognitive process of a radiologist, MuSiC-ViT was trained with multi-task learning and classified normal/abnormal, change/no change, and matching anatomy. A total of 406K CXRs were examined, with 88K pairs with changes and 115K pairs without changes recorded for the training dataset. For the internal validation dataset, 1,620 pairs were used. To show the robustness of MuSiC-ViT, we checked it with two external validation datasets. MuSiC-ViT achieved accuracy and area under the receiver operating characteristic curve (AUC) of 0.728 and 0.797 for the internal validation dataset, 0.614 and 0.784 for the first external validation dataset, and 0.745 and 0.858 for the second external validation dataset, respectively. In summary, we proposed a MuSiC-ViT that may discriminate between change and no-change CXR pairs by comparing baseline and follow-up CXR. By adding AMM, we proved through an ablation study that AMM helps AUC gain. Furthermore, disease loss to distinguish normal and normal of each baseline and follow-up CXR could help the model classify abnormal and normal CXRs in the case of disease. This architecture could be used to develop further CXR follow-up studies and lead to actual applications in clinical settings.