흉부방사선 영상에서 이상탐지에 강인한 정상군 선별을 위한 딥러닝 기반 다중학습 앙상블 모델 개발

Abstract

One of the primary responsibilities of radiologists in routine clinical settings is to diagnose chest radiographs (CXR) to determine the presence or absence of disease in patients. In scenarios where a large volume of chest radiographs are taken daily, the majority of patients presenting are diagnosed as normal. In such cases, the time and effort required for specialists to interpret every image are substantial, and the use of artificial intelligence models to classify definitively normal groups can significantly reduce the workload of doctors. However, the risk of missing significant diseases in patients is a critical issue. Therefore, an algorithm that accurately classifies normal and diseased states is necessary, especially one that avoids misdiagnosing diseased groups as normal. To address this, we propose the Strictly Chest X- ray Normal Network (SCXNN), a model that emulates the interpretation of radiologists and is more robust to diseases. The SCXNN model is a multi-task learning model that simultaneously performs reconstruction and classification for 13 different diseases. The first decoder in this model aims to learn the location and characteristics of lesions, enabling the encoder to learn the patterns of the disease. The second classifier, a Support Vector Machine (SVM), uses the logits from each disease model to classify groups into definite normal and abnormal. For data, we used 24,714 chest X-ray images collected from Asan Medical Center (AMC) in Seoul from 2011 to 2018, of which 5,400 were images with masks (Hard Label) and 19,314 were images without masks (Weak Label). The Hard Label data was split into training, validation, and test sets in an 8:1:1 ratio, and all Weak Label images were used for training. These images included or were classified as 'Normal' if they did not contain any of the 13 diseases such as Cardiomegaly, Advanced Tuberculosis, etc. Additionally, 355 chest X-ray images collected from January to March 2021 at AMC were used as external data for temporal validation datasets, and all mask and class labeling was performed by radiologists. Recognizing the problem of not sufficiently learning the characteristics of normal lung tissue when targeting only the diseased areas in the disease reconstruction process, we applied dilation, a type of morphological operation, in the mask processing step. This approach allowed for learning not only the diseased areas in the image but also the patterns of surrounding normal tissues affected by the disease, contributing to improved accuracy in disease reconstruction. The SCXNN model showed consistent performance in both internal and temporal validation datasets as well as external public datasets. This demonstrates the robustness of SCXNN, providing stable results regardless of the dataset type, which is crucial for ensuring patient health and safety in real clinical settings. Furthermore, the use of SCXNN as a diagnostic support tool for radiologists can contribute to reducing the workload of medical staff. This will enhance the speed and accuracy of the diagnostic process, improving the quality of medical services and allowing medical staff to focus more time on patient care, thereby enhancing the overall efficiency of medical services. The characteristics of SCXNN are expected to contribute to the further expansion of the use of AI tools in the clinical decision-making process.