Supervised and unsupervised learning for 3D brain image segmentation

M.Phil.Supervised and unsupervised automatic brain structure segmentation play pivotal roles in diagnosing brain diseases, including brain tumors, neuropsychiatric disorders and so on, thus are important in the clinical practice. However, developing an automatic approach to segment the brain structures remains very challenging in both supervised and unsupervised learning area. Supervised methods suffer from capturing the detail of complex brain structures, i.e., the ambiguous boundary, complex anatomical structures, and large variance in shapes of brain tissues. While for unsupervised segmentation methods, current registration-based segmenting architectures often fail to capture the hidden mapping relationship from unaligned image to reference image, thereby reducing the discriminativeness of learned features. In this thesis, to alleviate the challenging problems, we propose two deep-learning-based methods for supervised and unsupervised learning, respectively.First of all, we focus on the supervised learning method for 3D brain structure segmentation. We present a novel deep convolutional neural network ( -Net), which aims at boosting the information flow and enhancing the feature propagation. Specifically, we formulate a densely convolutional LSTM module (DCLSTM) to selectively aggregate the convolutional features (with the same spatial resolution) at the same stage of a CNN, which promotes the discriminativeness of features at each CNN stage. By stacking multiple DC-LSTMs among different stages, we can progressively and selectively aggregate the highly semantic features in deep stages to refine the fine detail features in shallow stages for better brain segmentation. Extensive experiments on two public benchmark datasets on sub-cortical brain structure segmentation validates the efficacy and effectiveness of our proposed network.Second, towards unsupervised segmentation, we study segmentation based on registration architecture, which transfers one brain image into the coordinate system of another brain image with the matched imaging contents. By employing this transformation relation between unlabeled images (unaligned images) and one brain volume with annotation (reference image), we can easily transfer the segmentation mask of the reference image to all unlabeled brain images. Specifically, to capture the internal transformation relation between a pair of 3D brain images, we develop a novel probabilistic multilayer regularization network for unsupervised brain image registration and segmentation. We first introduce direct regularizations into the hidden layers of two deep convolutional neural networks (CNNs) by adopting probabilistic models between hidden layers in these two CNNs. Furthermore, we embed the regularization terms into multiple layers of the CNNs and produce the feature-level latent variables in different layers. Combining the predicted feature-level latent variables of all layers, we can obtain a transformation metrics with rich information for a robust 3D brain image registration and segmentation. We employ two common benchmark datasets for unsupervised 3D brain image segmentation. Experimental results show that our method clearly outperforms state-of-the-art methods on both benchmark datasets by a large margin.監督式和無監督式的自動腦結構分割在診斷腦疾病的過程中起著關鍵作用(例如腦腫瘤, 神經精神疾病等), 因此該類方法在臨床實踐中具有重要意義。然而, 在監督式和無監督式學習領域中, 開發一種自動分割大腦結構的方法仍然非常具有挑戰性。監督式學習的方法難以捕獲复雜的大腦結構的細節, 例如腦溝模糊的邊界, 復雜的解剖結構以及大腦組織形狀的巨大差异。而當前基於配準架構的無監督分割模型通常無法捕獲從未對齊圖像到參考圖像的隱藏映射關系, 從而降低了用於圖像配準和圖像分割的特徵的判别能力。为了解决以上問題, 本文提出了兩種基於深度學習的監督式和無監督式的學習方法。首先, 針對於3D 腦結構分割的監督學習方法, 我們提出了一種新穎的深度卷積神經網絡( -Net) 來促進神經網絡的信息流和特徵傳播。具體來说, 我們設計了密集卷積長短期記憶模塊(DC-LSTM), 該模塊在CNN 的同一階段選擇性地聚合具有相同的空間分辨率的卷積特徵從, 而達到促進了每個CNN階段聚合特徵的判别能力。基於此, 通過在不同CNN 階段之間堆砌多個DC-LSTM, 我們可以逐步有選擇地聚合深層CNN階段的高度語義特徵, 用聚合的特徵來精煉更具精細細節的淺層CNN 階段的特徵, 以此實現更好的腦部分割。我們在兩個關於大腦皮層下結構的公共基準數據集上進行的大量實驗, 實驗結果驗證了我們提出的網絡的有效性和高效性。其次, 對於無監督式的分割方法, 我們研究方向主要針對基於配準架構的分割方法。該類分割方法主要涉及到將一個大腦圖像轉移到另一個大腦圖像的坐標系的配準的過程。通過應用配準過程中的没有分割結果的圖像(未對齊圖像) 和帶有分割結果的圖像(參考圖像) 間的轉化關系, 我們可以輕鬆地將參考圖像的分割結果轉化到所有没有分割結果的大腦圖像上。具體來说, 为了捕獲一對3D 腦圖像之間的内部轉化關系, 我們提出了一種新穎的多層概率正則化模型來對腦圖像進行無監督配準和分割。。我們首先在兩個深度卷積神經網絡的隱藏層之間建立概率模型, 將概率正則化模型直接加入這兩個深度卷積神經網絡的隱藏層。此外, 我們將概率正則化模型項嵌入到CNN的多個階段中, 然後在不同CNN 階段中生成一個特徵級的隱藏變量。結合所有階層預測出來的特徵級的隱藏變量, 我們可以獲得一個具有更豐富轉化信息的轉化指標, 基於此指標來達到更爲魯棒的3D 腦圖像的無監督式配準和分割。我們在兩個常見的基準數據集進行無監督的3D 腦圖像配準和分割。實驗結果表明, 在兩個數據集上, 我們的方法明顯優於最新方法。Liu, Lihao."December 2019."Thesis M.Phil. Chinese University of Hong Kong 2020.Includes bibliographical references (leaves 74-85).Abstracts also in Chinese.Title from PDF title page (viewed on February 17, 2022)