隨著深度學習在影像處理領域的發展,有越來越多研究者開始以深度學習技術應用於醫學影像分析,在此領域中影像分割是一個常見的議題,如從圖像中找到精確的器官、腫瘤或血管等等,這些分割結果可能會直接應用於最後的結果 (eg. 評估大小),或是作為後續分類、計算分數的前置資料。 在影像分析演算法開發與部署的過程中,會隨著案例不同而有各自的問題需要處理,在演算法開發上,我們以微循環影片分析做為案例,因為微循環影像的複雜度導致血管標註工作需要耗費大量人力,我們嘗試使用傳統電腦視覺方法生成的標註輔以深度學習模型強大的泛化能力來完成血管分割的任務;而在演算法部署上,我們以心血管鈣化分數做為案例,因為演算法的處理流程中會有耗時的後處理,導致使用 PyTorch For-Loop 推論架構會有大量時間的資源閒置,我們嘗試設計一個事件驅動的架構來處理。 在最後成果上,在微循環影片分析上,我們發現以 SATO 血管分割演算法生成的標註結合醫學影像常使用的 UNet 可以捕捉到比原先生成的標註更多的血管,展示了以電腦視覺方法生成的標註可以訓練出更優秀的深度學習模型的潛力;而在心血管鈣化分數計算上,事件驅動的架構可以顯著提升整體推論速度,同時也成功將基於 HeAortaNet 的心血管鈣化分數演算法應用於健保醫學影像資料庫。
In recent years, many researchers have begun to apply deep learning technology to medical image analysis, because of the rapid development of deep learning in the field of image processing. Image segmentation tasks such as finding precise organs, tumors or blood vessels from images is a common topic. These segmentation results may be used directly (eg. evaluation the organ size), or as the preliminary data of subsequent classification, calculation of scores. There are different problems that need to be dealt with depending on the case in the image analysis algorithm development and deployment. In algorithm development, we take the micro-circulation video analysis as an example. We try to use labels generated by the traditional computer vision (CV) methods and the generalization ability of deep learning (DL) model to complete the blood vessel segmentation. This will solve the problem that the blood vessel labeling being too complex to label. In the algorithm deployment, we take the cardiovascular calcification score algorithm as an example. We design an event-driven architecture to reduce the resource idling. Because of the time-costumed post-processing will block in PyTorch For-Loop architecture. Finally, in the micro-circulation video analysis development, we shows the labels generate by CV methods is possible to train a better DL model. Because of the UNet be trained with the labels generated by the SATO vessel segmentation algorithm can capture more vessels than the original labels. On the other hand, in the cardiovascular calcification score algorithm deployment, event-driven architecture can significantly improve inference speed over For-Loop architecture. We have also successfully applied the cardiovascular calcification score algorithm based on HeAortaNet to the medical imaging data of the National Health Insurance Agency.