結合機器學習與Dixon技術來自動化頭頸部肌肉脂肪定量分析

越來越多研究指出咬肌是影響頭頸癌患者治療及預後情形的關鍵因素，隨著年紀增長，脂肪在肌肉中不斷累積，讓衰老的肌肉功能下降，引發肌少症等年老問題。近年來Dixon 方法透過後處理方式顯示出水與脂肪分離影像，可用於量化脂肪含量的脂肪分率（fat fraction）影像。本研究旨在運用機器學習技術與磁振造影（MRI）影像資料，發展一套自動化分析流程，以探討頭頸癌（Head and Neck Cancer, HNC）患者之頭頸部相關肌肉的脂肪浸潤狀況，並進一步評估其與臨床特徵及治療成效的關聯性。研究對象包含正常對照組與頭頸癌病人組，共61名受試者，其中病人組34人，正常組27人，另有19位病人同時具備化療前後的影像資料。每位受試者皆接受T2 Dixon MRI掃描，以獲得水脂分離的影像資訊，用以計算肌肉脂肪分率。影像資料首先經由深度學習模型進行肌肉區域的自動分割。本研究選擇三塊具代表性的肌肉區域，包括咬肌（masseter muscle, MM）、內翼肌（medial pterygoid muscle, MPM）與外翼肌（lateral pterygoid muscle, LPM）。為提升分割精度，採用多種深度學習架構進行比較，包括U-Net、U-Net++、3D U-Net以及3D Res-U-Net等模型，並針對影像中常見的小肌肉與黏附肌群問題引入殘差機制。分割模型以Dice係數、Jaccard index與Conformity index進行性能評估，結果顯示U-Net++於整體表現最佳。完成分割後，計算各肌肉區域的脂肪分率，並進行群體間統計比較。統計分析結果顯示，病人組肌肉脂肪分率普遍高於正常組，且在咬肌的區辨能力上最為顯著（AUC約0.78）。此外，病人組的脂肪分率分布並不呈常態，與不同因子具有顯著相關性。治療前後的比較亦顯示三塊肌肉脂肪含量在化療後顯著改善，證明脂肪分率可作為治療反應的潛在影像生物標記。

Recent studies have highlighted the critical role of masticatory muscles, particularly the masseter, in influencing treatment outcomes and prognosis in patients with head and neck cancer (HNC). Aging is associated with increased fat infiltration in skeletal muscles, contributing to functional decline and conditions such as sarcopenia. The Dixon MRI technique enables water–fat separation through post-processing, allowing for the quantification of muscle fat content via fat fraction imaging.This study aims to develop an automated analysis pipeline integrating machine learning techniques and magnetic resonance imaging (MRI) data to assess fat infiltration in head and neck muscles of HNC patients, and to investigate its associations with clinical characteristics and treatment response.A total of 61 participants were enrolled, including 34 HNC patients and 27 healthy controls. Among the patients, 19 had paired MRI scans acquired before and after chemotherapy. All subjects underwent T2-weighted Dixon MRI to obtain water–fat separated images for fat fraction calculation. A deep learning-based segmentation approach was employed to automatically delineate three representative muscles: the masseter muscle (MM), medial pterygoid muscle (MPM), and lateral pterygoid muscle (LPM). To improve segmentation accuracy, multiple deep learning architectures were compared, including U-Net, U-Net++, 3D U-Net, and 3D Res-U-Net. A residual learning mechanism was introduced to address segmentation challenges posed by small and closely attached muscle groups. Model performance was evaluated using the Dice coefficient, Jaccard index, and Conformity index, with U-Net++ demonstrating the highest overall accuracy.Following segmentation, fat fractions of each muscle were quantified and statistically analyzed. Results indicated significantly higher fat fractions in the patient group compared to controls, with the masseter muscle exhibiting the strongest discriminatory power (AUC ≈ 0.78). Fat fraction distributions in the patient group were non-normally distributed and showed significant correlations with various clinical parameters. Additionally, paired comparisons revealed a significant reduction in fat content across all three muscles after chemotherapy, supporting fat fraction as a potential imaging biomarker for treatment response in HNC.