DIXON核磁共振影像之自動分割及放射學研究及其骨骼肌與肝臟組織之應用

Abstract

Radiomics是近年來放射科醫學領域中新興的研究方法，首先分割影像中的感興趣區域（region of interest，簡稱ROI），進而提取圖像ROI中的數據特徵，以分析特徵與疾病間的關聯性，最後建立疾病判別模型以協助臨床診斷、預測以及預後。 Radiomics研究過程中的影像ROI圈選非常重要，圈選誤差可能影響提取之特徵值，進而降低分析結果的正確性，若採用手動圈選ROI將費時費力，且過程中將受到諸多外在因素影響，導致結果不穩定。當ROI為複雜形狀時，圈選過程將更加困難，於部分研究中也提出利用自動或半自動軟體進行ROI分割。完成影像分割後即可進行特徵提取，可提取的特徵其實可以有無限多種，但即便建構一有效預測疾病之模型，也可能因模型中過度複雜的特徵定義，或是因較難描述之影像信號意義，而限制其特徵解釋力。目前文獻上很少有研究提出一套完整包含影像自動分割、特徵提取及特徵分析解釋之整合研究，因此本研究期望整合影像分割、特徵萃取、特徵分析之完整研究流程，並利用特徵解釋組織與疾病關聯性，藉此強化研究結果之再現性及可信度。 近年來深度學習於醫學圖像分割的發展已逐漸成熟，針對不同的資料格式及複雜的ROI形狀皆有因應的模型架構，本研究於圖像分割的過程中針對2D、3D影像不同的特性分別採用「加強全卷積模型」以及「連續影像全卷積模型」，進一步改良模型架構，提升圈選效能，並評估兩模型採用遷移學習後的差異，期以自動、準確且客觀的圖像ROI判別，使後續的特徵分析更加精確。而為了使Radiomics特徵解釋更加直觀，本研究應用DIXON MRI中純水、純脂肪信號之影像作為研究對象。 驗證本研究於2D、3D DIXON MRI醫學圖像之Radiomics研究流程，分析兩筆實際案例－骨骼肌（2D影像）、肝臟組織（3D影像），而應用之疾病分別為多囊腎、脂肪肝。研究結果顯示，本研究成功改良2D影像應用之圈選模型，取得較原先模型更佳腹部骨骼肌的圈選效能，並利用骨骼肌之水、脂肪訊號的散佈圖及直方圖特徵萃取，了解多囊腎於骨骼肌成份之影像。本研究在有限的3D影像樣本中，亦證實透過連續影像全卷積模型可以準確分割3D肝臟組織影像，並萃取水、脂肪訊號的直方圖特徵以二分法將脂肪肝等級分為「0 vs.1,2」建立羅吉斯模型區分脂肪肝等級，可於測試資料中區分「0 vs.1,2」及「0,1 vs. 2」病變等級之AUC分別達到0.917、1，表示模型對於脂肪肝等級區分具有一定的判別能力。此外，本研究應用線性回歸模型預測測試資料的病變等級，估計值明顯區分出各等級之樣本，並以肝臟脂肪含量比例建構線性迴歸模型，於僅僅25筆訓練資料中，模型R^2可達0.88，且於測試資料中的應變數及估計值之相關係數達0.93。上述各項驗證結果，證實了本研究具有一定的有效性，本研究最後也透過選取之特徵發掘了組織特徵與疾病的潛在關聯性。

Radiomics is a new trend of research method in Radiology recently. The research process of Radiomics includes segmentation of the region of interest (ROI), extraction of quantitative features resulting from the conversion of images data, and the subsequent analysis of these features for clinical decision support. Segmentation plays a vital role in the Radiomics research process because its error may affect the extracted feature values, thereby diminishing the correctness of the analysis results. However, manual segmentation is subjective, error prone and time consuming, especially when ROI is a complex shape. In some studies, existing automatic and semi-automatic software are used for ROI segmentation. After the ROI segmentation, feature extraction can be performed and there are numerous kinds of features that can be created. But even if an effective model for predicting a disease is constructed, the model interpretation is still limited by its complex feature definitions or the relatively indescribable source of image signals. In this study, we integrate the complete process of Radiomic studies, including ROI segmentation, feature extraction, feature analysis, and the association between tissue characterization and the disease by the extracted features. Moreover, we attempt to improve the reproducibility and credibility of the models. Deep learning-based image segmentation is by now firmly established as a robust tool in image segmentation. For different data formats and complex ROI shapes, there exist corresponding useful deep-learning model architectures. For this reason, we will adopt 'Fully Convolutional Networks with Image Preprocessing' and 'Consecutive Fully Convolutional Networks' described in the literature to segment ROI for 2D and 3D image, respectively. In order to improve the accuracy and stability of model segmentation, we modified the 2D model architecture and compare the results with and without Transfer Learning. In this study, we focus on the DIXON MR images and make interpretation of Radiomics features more intuitive by considering the pure water and pure fat signals. In order to verify the Radiomic studies of 2D and 3D DIXON MRI, two actual cases were analyzed: 2D DIXON images of skeletal muscle (SM) and 3D DIXON images of liver tissue. For the SM images, we use the cases collected for patients with the polycystic kidney disease. For the liver 3D images, we use the cases collected for patients with liver steatosis. This study results show that the segmentation model of 2D image has a better performance than the original model in segmenting the ROI of SM. And through the SM feature extraction of the scatter plot and histogram of water and fat signals, we demonstrate how to interpret the effect of polycystic kidney on the skeletal muscle tissue. This study also confirms that the 'Consecutive Fully Convolutional Network' can accurately segment 3D liver tissue images in a limited sample of DIXON images. Using the features extracted from the histogram of water and fat signals, we establish Logistic regression models to differentiate liver steatosis grade 0 from grades 1 and 2 and grade 2 from grades 0 and 1 with testing AUC effectively reaching 0.917 and 1, respectively. In addition, a linear regression model is established to estimate the fat percentage. Results show that the R2 of merely 25 training cases reaches 0.88 and the correlation of real and estimated fat percentage in testing cases can reach 0.93.