利用多尺度3D U-Net在電腦斷層掃描影像上進行腎上腺的自動化分割

Abstract

腎上腺是人體重要的內分泌器官，分泌過少或過多的激素都可能會引發嚴重的疾病。醫學影像技術，例如電腦斷層掃描（Computed Tomography，CT），因其非侵入性的特點而常被用於診斷腎上腺問題。醫學影像分割可以幫助深入瞭解影像，提取特徵並評估異常或病灶。醫師透過醫學影像解讀可以分析腎上腺的位置和特徵，有助於診斷和治療計劃的制訂。然而，在電腦軟體手動標記腎上腺的位置可能非常耗時，加上每天有大量臨床工作，這無疑造成了醫師龐大的工作負擔。 現今，深度學習由於其高準確性和時間效率，在醫學影像分割上被廣泛應用。但深度學習仍然存在一些困難，特別是在對像腎上腺這種小而形狀不規則的器官進行分割時，模型可能會難以提取特徵。基於上述，我們建立了一個三維U-Net模型，結合了多尺度和集成學習作為訓練方法，來對醫學影像進行腎上腺的自動化分割。本研究以腹部CT影像作為訓練資料集，除了將影像資料的Z軸涵蓋範圍調整成統一的大小之外，不做其他影像前處理，藉以驗證深度學習模型從CT影像中識別和學習腎上腺結構特徵的能力。模型的分割能力使用Dice係數進行評估。 實驗結果顯示，將三維U-Net與多尺度和集成學習結合使用，在腎上腺的分割中可達到66.3%的準確度，值得進一步驗證臨床實用性。

Adrenal glands are vital endocrine organs in the human body, playing a crucial role in hormone production. Imbalances in these hormones can result in serious disorders. With the advantage of minimal noninvasiveness, medical imaging techniques, such as x-ray computed tomography (CT), are commonly used to diagnose adrenal gland disorders. Image segmentation enables feature extraction and evaluation of lesions/abnormalities, and whereby helps diagnosis and treatment planning. However, manual marking of adrenal glands positions on a computer can be time-consuming, posing a significant workload for doctors. Nowadays, deep learning has been widely used for segmentation in medical images due to its high accuracy and time efficiency. However, there are still challenges to overcome, particularly in segmenting small and irregularly shaped organs like adrenal glands in medical images. In light of the above, we built a three-dimensional U-Net in combination with multi-scale and ensemble learning to automatically segment and label adrenal glands on CT images. Except for resizing the tensor size along the z-axis of the data, no other image preprocessing was performed so as to validate the ability of deep learning in identifying and learning the features of adrenal structures from CT images. The segmentation performance of the model was evaluated using the Dice coefficient scores. According to our results, 3D U-Net combined with multi-scale and ensemble learning achieved an average accuracy of 66.3% in adrenal gland segmentation. The proposed method warrants further investigation for its clinical application.