基於 3D 單光子電腦斷層心肌灌流影像的冠狀動脈疾病預測深度學習技術

Abstract

本研究主要探討了二維和三維單光子發射計算機斷層攝影（SPECT）心肌灌注影像（MPI），以及深度學習的卷積神經網絡在冠狀動脈疾病（CAD）檢測上的應用。我們的目標是利用這些技術提升醫生進行視覺診斷的準確率。由於深度學習在醫學影像分類方面已取得重要的突破，我們在此研究中採用了多模態模型融合 (Fusion) 來進一步優化對冠狀動脈疾病的識別。 除了從原始的三維 SPECT MPI 影像中提取空間資訊，我們還結合二維極地圖 (Polar Map)和臨床數據進行模型訓練。本研究評估了不同影像預處理技術對模型性能的影響，並探索了多種融合模型方法，將不同類型影像的深度學習模型，以提升診斷準確性的可能性。 根據最新的研究結果，我們在深度學習模型中納入臨床數據後，模型的表現得到了提升。此外，我們還發現將不同類型的模型以串接方式進行融合可以達到最佳效果。本研究還利用序列式多門檻設定進行疾病分類，在冠狀動脈疾病的識別方面，我們開發的深度學習模型利用 SPECT-MPI 和臨床數據的結合，能夠更準確地進行診斷，其 AUC 值達到了 70% 以上。這一效果優於專家的臨床診斷能力，因此有望應用於冠心病的診斷中。 總結而言，這項基於深度學習方法的冠心病診斷研究具有巨大的應用潛力。根據我們的實驗結果，這項研究成果可以作為臨床決策支援系統中的工具，為醫生提供有效的支持，提高冠心病診斷的準確性和效率。透過整合臨床數據和深度學習模型，醫生將能夠更快速且更準確地進行診斷，並提供更適切的治療方案。

This study primarily investigated the application of two-dimensional and three-dimensional single-photon emission computed tomography myocardial perfusion imaging (SPECT MPI) and deep learning convolutional neural networks in the detection of coronary artery disease (CAD). Our goal was to enhance the accuracy of visual diagnosis for physicians using these techniques. Considering the significant breakthroughs of deep learning in medical image classification, we employed multimodal model fusion in this study to further optimize the identification of coronary artery disease. In addition to extracting spatial information from the original three-dimensional SPECT MPI images, we also incorporated polar maps and clinical data for model training. The study evaluated the impact of different image preprocessing techniques on model performance and explored various fusion methods to improve diagnostic accuracy through the combination of deep learning models for different types of images. Based on the latest research findings, we observed an improvement in model performance when clinical data was integrated into the deep learning model. Furthermore, we discovered that fusing different types of models through concatenation yielded the best results. The study also employed a sequential multi-threshold approach for disease classification, achieving more accurate diagnosis of coronary artery disease by leveraging the combination of SPECT MPI and clinical data, with an AUC exceeding 70%. This outperformed the clinical diagnostic capabilities of experts in the early stage, suggesting promising applications in coronary heart disease diagnosis. In conclusion, this study on deep learning methods for coronary artery disease diagnosis has significant potential for practical applications. Based on our experimental results, the findings can be integrated into clinical decision support systems as a tool to provide effective support for physicians and improve the accuracy and efficiency of coronary artery disease diagnosis. By integrating clinical data with deep learning models, physicians will be able to diagnose more rapidly and accurately, leading to more appropriate treatment strategies.