| 摘要: | 減骨圖像處理技術常用於斷層掃描血管造影 (CTA) 以協助臨床診斷，近年來深度學習技術的蓬勃發展，通過採用由圖形處理器 (GPU) 實現的深度卷積神經網絡模型可以進一步改進減骨 CTA。監督式學習已被證明是有效訓練人工智能分析醫學圖像的方法，其中標記數據集的品質在模型收斂速度中扮演至關重要的作用。然而，大多數現有方法在標記醫學圖像方面的一個主要弱點是需要領域專家知識。此外，標記“大數據”所需的極高時間成本使得三維 (3D) 醫學數據標記成為一項極具挑戰性的任務。由於合格的專家人數有限和 3D 數據註釋動脈和靜脈血管的困難度高。標記 CTA 血管在現實環境下難以實現。相比之下，標記CTA骨骼是一種更可行的方法。在這項研究中，我們提出了一種圓柱採樣策略，以幫助非專家標記下頜骨和脊柱以進行電腦斷層掃描 (CT)的標記與機器學習訓練。它利用了相對於身體縱軸的準解剖對稱性。這種圓柱採樣方法允許生成帶有泛解剖地標的 2D 採樣掃描，這些解剖地標是透過沿縱軸有較佳的解剖結構連續性旋轉不同角度中得出的。本研究包含 20 個電腦斷層掃描血管造影數據集， 使用我們提議的圓柱採樣掃描方法以手動或自動標記C1、C2脊柱骨和下頜骨，並記錄和比較每個骨骼的標記時間。在圖像標記過程之後，我們使用以 FPN-ResNeXt 為主幹的深度學習語義分割Unet模型，將所提出的圓柱掃描與傳統（橫向、矢狀和冠狀）掃描做機器學習收斂性的比較分析。隨機選擇15個CTA數據（75%）作為訓練集，其餘5個（25%）作為測試集，實驗分別用 1、5、10 和 15 個CTA數據訓練模型，每個實驗分別重複5次。總體而言，該神經網絡模型在圓柱、橫向、矢狀和冠狀掃描中分別未能收斂 1 (5%)、12 (60%)、1 (5%) 和 5 (25%) 次。圓柱採樣方法在 C1、C2 和下頜骨分割中分別產生了 91.3%、92.8% 和 93.8% 的最高測試 F1 分數。用於標記下頜骨的時間分別為，每個圓柱採樣的CTA數據約需要 90 分鐘而每個橫向採樣的CTA數據則約需要190分鐘。實驗結果顯示，我們所提出的電腦斷層掃描圓柱採樣方法不僅減少了標記所需時間並且可以訓練出更準確的骨骼分割深度學習神經網絡模型。;Bone-subtraction is an image processing technique that is often used in clinical settings to enhance computed tomography angiography (CTA) interpretation. With the newly developed deep learning technology, bone-subtraction CTA can potentially be further improved by adopting deep convolutional neural network (CNN) models implemented with advanced graphic processing units (GPU). Supervised learning has been proven to be an effective approach to creating artificial intelligence for medical image analysis, where the quality of labeled datasets plays the essential role in the speed of model convergence. However, a major drawback of most existing approaches in labeling medical images is that domain expert knowledge is required. Further, the extremely high time cost for labeling “big data” has made three dimensional (3D) medical data labeling a challenging task. Labeling vessels in CTA is difficult to achieve in practice due to shortage of qualified experts and the intrinsic difficulty in annotating arterial and venous vessels in 3D data. In contrast, labeling bones is a more feasible approach. In this study, we propose a cylindrical sampling strategy to assist non-experts in labeling mandible and spine for computed tomography (CT) scans. It takes advantage of the quasi-anatomical symmetry in respective to the body’s longitudinal axis. This cylindrical sampling approach allows generating 2D resampled scans with pan anatomical landmarks which are derived from the continuum of anatomical structures by rotating different angles along the longitudinal axis. Twenty computed tomography angiography datasets were included in this study. C1, C2 spinal bones, and mandibles were manually or automatically labeled by using the proposed cylindrical resampling scans. The labeling time for each bone was recorded and compared. After the image labeling process, we also conducted an experiment in comparing the convergency of the proposed cylindrical scans with traditional (transverse, sagittal, and coronal) scans using a deep learning semantic segmentation Unet model with an FPN-ResNeXt backbone. Fifteen subjects (75%) were randomly selected as the training set, and the remaining 5 subjects (25%) were used as the test set. The model was trained with 1, 5, 10, and 15 subjects respectively. The experiment was repeated for 5 times. In total, the model failed to converge 1 (5%), 12 (60%), 1 (5%), and 5 (25%) times for cylindrical, transverse, sagittal, and coronal scans respectively. The cylindrical approach generated the highest test F1-scores of 91.3%, 92.8%, and 93.8% in C1, C2, and mandible segmentation respectively. The time used in labeling cylindrical and transverse mandible scans was 90 and 190 minutes per subject respectively. The experimental results show that the proposed cylindrical sampling method for head and neck CTA not only reduces the labeling time but also achieves better segmentation of bones. |
