다중 테스크 학습을 통한 DWI-FLAIR 영상의 신호 불일치를 이용한 뇌졸중 onset 예측에 관한 연구: 외부 데이터셋들을 이용한 모델 검증

Abstract

The treatment of acute ischemic stroke, which has a strong correlation between the time clock and tissue progression, heavily relies on the onset time. However, determining the exact timing of onset can be challenging due to factors such as stroke occurrence during sleep or unclear onset. To address this issue, previous studies have utilized the signal mismatch between DWI and FLAIR images to estimate the elapsed time after stroke onset. In this study, our goal is to develop a technique for determining the time elapsed after stroke using deep learning with multi-task learning (MTL) approach based on MRI image biomarkers. We aim to evaluate the performance of this technique using external validation datasets. The proposed model takes a combination of ADC, FLAIR, and DWI-b1000 volumes as input, and performs simultaneous prediction of stroke onset time and infarct segmentation for time windows of 3 hours, 4.5 hours, and 6 hours using MTL. The backbone of the model is the 3D patch based SwinUNETR model, with the addition of auxiliary classifiers in the bottleneck layer. A voting ensemble of patch-level classifiers is performed at the patient level for classification. The auxiliary classifiers count and vote for onset identification based on the presence of infarction in patches using the predicted mask. For the classification task, we compared the performance of our model with 3D DenseNet and 3D patched DenseNet using internal validation, using AUC, accuracy, specificity, and sensitivity as evaluation metrics. We also compared the performance of onset classification for 4.5 hours using two external validations. For the segmentation task, we compared the performance of DiNTS, UNETR, nnUNET, and our proposed model using IOU and Dice coefficient as evaluation metrics. Overall, our proposed model showed superior performance in classification compared to the compared models, and it also demonstrated similar or improved segmentation results compared to nnUNET. Additionally, when predicting onset time, the model showed attention to the lesion by extracting Grad-CAM from the encoder, indicating its focus on the infarction during training. We provided both quantitative and qualitative evaluations across multiple segmentation and classification tasks. This MTL-based model showed better performance in identifying stroke onset time in internal and two external validations, demonstrating its potential for potential clinical use. The proposed model has the potential to assist in performing thrombolysis therapy at the appropriate time for stroke patients with unclear onset time by predicting the onset time accurately.