2.5D U-Net 級聯式深度學習框架應用於電腦斷層圖像中肝臟及腫瘤自動分割

Abstract

電腦斷層是目前肝臟腫瘤診斷中應用最廣泛的醫學成像方式。傳統上，放射科醫師透過肉眼對電腦斷層圖像中別出肝臟和肝臟的腫瘤區域後，再以逐一切片勾畫肝臟和腫瘤的方式，作為後續治療的依據。然而，該工作相當耗費時間與勞力，而且缺乏一個客觀且明確定義的判斷方法。因此，在臨床實踐中，肝臟和腫瘤分割過程的自動化是有價值的。 近年來，深度學習的技術促使醫療影像語義分割的結果有了顯著的進步，其中 3D 分割網路在許多的相關任務當中取得優異的表現。然而，3D 卷積本身有計算資源上的限制，2D 卷積則難以學習到相鄰 CT 切片之間的上下文資訊。因此，本論文提出了一個基於 2D 的深度學習框架用於電腦斷層圖像中肝臟及腫瘤的自動分割，並應用一些相關文獻提出的可以提升 2D 分割網路表現的方式，包含輸入多張相鄰的二維切片，兩階段級聯式的模型建置方式，使用 EfficientNet 做為分割網路的編碼器，以及三平面集成。如此，所提出的方法能夠在相對較低的計算資源需求之下達到良好的肝臟和腫瘤分割表現。 除此之外，多數相關研究只注重在整個模型建置流程其中一個層面上的改良，例如網路架構或是損失函數的設計。故本論文將以整個深度學習框架當中的各項實作細節作為研究主體，從一開始的資料前處理、網路架構、損失函數、多模型分割結果集成，再到最後的後處理，試圖從相關實驗中找出一個最佳的方法組合以更進一步地提升所提出方法肝臟和腫瘤的分割準確度。 本篇論文的研究貢獻如下。第一，本論文提出了一個基於 2D 的深度學習框架用於電腦斷層圖像中肝臟及腫瘤的自動分割，以解決相關 3D 方法於計算資源上的限制，其中輸入多張相鄰的二維切片，兩階段級聯式的模型建置方式，使用 EfficientNet 做為分割網路的編碼器，以及三平面集成的作法使其能夠在相對較低的計算資源需求之下達到良好的肝臟和腫瘤分割表現。第二，多數相關研究只注重在整個模型建置流程其中一個層面上的改良，故本論文以整個深度學習框架當中的各項實作細節作為研究主體，從一開始的資料前處理到最後的後處理。基於從實驗結果得到的最佳方法組合，所提出方法於肝臟和腫瘤分割的表現有更進一步的提升，在 LiTS 測試集肝臟分割的 Dice per case 達到 0.9660 的水準，腫瘤分割的 Dice per case 則達到 0.7180 的水準。同時，這些實驗結果將會形成一套分割模型表現的改善方案，可供從事相關工作的研究人員參考。

Computed tomography is currently the most widely used medical imaging method in liver tumor diagnosis. Traditionally, radiologists identify liver and tumor regions from CT images with naked eyes and then delineate these regions in a slice-by-slice manner as the basis for subsequent treatment. However, this work is time-consuming and labor-intensive, lacking an objective and clearly defined measurement method. Therefore, in clinical practice, the automation of the liver and tumor segmentation process is valuable. In recent years, deep learning techniques have led to significant progress in the results of semantic segmentation of medical images, in which 3D segmentation networks have achieved excellent performance in many related tasks. However, 3D convolution has limitations in computational resources, and 2D convolution is hard to learn the context information between adjacent CT slices. Therefore, this thesis proposes a 2D-based deep learning framework for automatic liver and tumor segmentation on computed tomography images and applies some approaches proposed in the related literature that can improve the performance of 2D segmentation networks, including inputting multiple adjacent 2D slices, two-stage cascaded model building approach, using EfficientNet as the segmentation network's encoder, and the triplanar ensemble. This way, the proposed method can achieve great liver and tumor segmentation performance with a relatively low computational resource requirement. Moreover, most related research only focuses on improving one aspect of the entire model building process, such as the design of network architecture or the loss function. Thus, this thesis will take the implementation details of the whole deep learning framework as the research subject, from the initial data preprocessing, network architecture, loss function, multi-model segmentation results ensemble, and to the final post-processing, trying to find the best method combination from the related experiments to improve the proposed method's liver and tumor segmentation accuracy even further. The research contributions of this thesis are as follows. First, this thesis proposes a 2D-based deep learning framework for automatic liver and tumor segmentation on computed tomography images to address the computational resource constraints of related 3D methods, in which the implementation of inputting multiple adjacent 2D slices, two-stage cascaded model building approach, using EfficientNet as the segmentation network's encoder, and the triplanar ensemble enable it to achieve great liver and tumor segmentation performance with a relatively low computational resource requirement. Second, most related research only focuses on improving one aspect of the entire model building process, so this thesis takes the implementation details of the whole deep learning framework as the research subject, from the initial data preprocessing to the final post-processing. Based on the best method combination obtained from the experimental results, the proposed method further improves its liver and tumor segmentation performance, achieving a Dice per case of 0.9660 for liver segmentation and a Dice per case of 0.7180 for tumor segmentation on the LiTS test set. Meanwhile, these experimental results will form a set of improvement schemes for the performance of the segmentation model, which can serve as a reference for researchers engaged in related works.