由於大數據及硬體計算能力的進步，人工智慧 (artificial intelligence，AI) 已廣泛滲透日常生活，其醫療應用亦日益增加。醫療影像需要放射科或病理科醫師的判讀，不僅耗費人力時間，判讀時也可能錯過超越人眼辨識範圍的細微病變。醫療影像數位化及圖像資料的累積，讓AI展現其價值，以AI方法進行影像判讀，可以有效提效率並減少失誤。全卷積神經網絡 (fully convolutional network，FCN) 是目前以AI處理數位醫學影像最常使用的方法，藉由監督式深度學習提取圖片中的特徵值，以此分類結果或分割異常區域。然而監督式的影像辨識分割有一個缺點，即必須要由領域專家們在圖片上做標記，把不良的組織標註出來，再輸入到深度學習中的神經網路做模型訓練。這些前置作業非常耗時，而且必須犧牲眾多專業人力。為避免調校標註耗時和發生標記錯誤的可能，本研究另行研究以密度泛函理論為依據建構之Data structures with Density Functional Theory (DDFT) 非監督式學習演算法，針對腦部核磁共振影像 (Magnetic Resonance Imaging，MRI)，進行不良組織的辨識判讀與分割。此外，該方法也將與由FCN延伸出的U-net進行比較，藉由正確率與效能評估是否能以DDFT非監督式學習輔助U-net監督式學習。;Methods of image segmentation of medical imageries always benefit clinical investigation, anatomic researches, and modern medicine. The high variances of tissue morphologies, however, obstruct and threaten the development and the feasibility of contemporary techniques. Compared to traditional methods of biomedical image processing, the deep-learning-based methods may offer an avenue to deal with the mentioned predicaments. Among these emerging techniques, the fully convolutional neural network (FCN) as well as the U-net are attracting attention from the relevant participants. In the framework of FCN, it is possible to accept input images of any size and it is also not necessary requiring that all training images and test images have the same sizes. Additionally, the FCN can avoid the problems of repeated storage and repeated convolutional calculations caused by the use of pixel blocks. However, the time consumption in training procedures and relevant lesion labeling is also a challenging issue in applications. To conquer the problems, we propose a hybridized algorithm by introducing the concept of levelset from image processing methods and the learning block from that of machine learning into the data density functional method. The learning block was embedded in the data density functional theory (DDFT) to balance the Lagrangian level, and the level set was then used to set these levels for image segmentation. Two brain Magnetic Resonance image sets were employed to demonstrate the validation of the proposed hybridized method. Eventually, the comparison of segmentation performance between the proposed algorithm and the U-net are discussed as well as the limitations of the proposed hybridized algorithm.
