利用機器學習量化評估經放射線治療後頭頸癌患者之肌肉超音波影像

Quantitative Evaluation of Muscle Ultrasound Images Using Machine Learning for Head/Neck Cancer Patients with Radiation Therapy

許晏寧(Yen-Ning Hsu)

指導教授 ： 林文澧

國立臺灣大學/工學院/醫學工程學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202002794

摘要

研究背景與目的
頭頸癌患者經放射線治療後，通常會發生肌肉纖維化的副作用，造成該肌肉僵硬與緊繃，使脖子無法轉動，嚴重地影響病患的生活品質，因此他們會定期且長期回醫院追蹤與檢查。在臨床診斷上，醫師通常利用觸診及肉眼判斷放射線治療前與後的肌肉超音波影像之變化情形，屬於主觀和定性的診斷方式。因此，本研究提出自動分割超音波影像上之肌肉與纖維分析的方法，輔助醫師以客觀和定量的角度評估肌肉受傷情形，進而預測放射線治療後的情況，藉此優化其治療策略與復健療程，改善病患往後的生活品質。
材料與方法
本研究共完整收案28位經66~70 Gy放射線治療的頭頸癌患者，包括20位男性和8位女性，年齡座落於28~69歲，採用其放射線治療前與後追蹤一年內的B型超音波影像。在影像分析方法上，主要分為肌肉分割與纖維量化這兩部分。首先，訓練U-Net深度神經網路來自動分割超音波影像上之胸鎖乳突肌，接著，將其肌肉進行Gabor濾波器和條件型自適應閾值演算法來概略分割纖維組織，此外，再利用單類別支持向量機涉及空間分布與表現，以提取更精確的纖維組織，最後，計算纖維比與其變化量以客觀評估該肌肉纖維化的嚴重程度。
結果
本研究結果顯示基於U-Net所分割超音波影像上之胸鎖乳突肌的評估指標，包括準確率達99%，Dice coefficient達96%，而精確率及召回率也達96%。在纖維分析方面，本研究發現放射線治療後3個月及6個月相較於放射線治療前的纖維比變化量有明顯增加，而1年後則有緩降趨勢。
結論
本研究提供一個客觀與量化的方式來分割超音波影像上之胸鎖乳突肌及分析其纖維比趨勢，而此方法可用於預測肌肉纖維化的預後情形，並輔助醫師找出最佳治療策略。

關鍵字

B型超音波影像 ； 頭頸部癌症 ； 肌肉纖維化 ； U-Net ； Gabor濾波器 ； 條件型自適應閾值演算法 ； 單類別支持向量機

並列摘要

Background and Objective
Patients with head/neck cancer after receiving radiation therapy (RT) usually suffer from the side effects of muscle fibrosis, causing muscle stiffness and tightness and restricting the neck movement, and these deeply affect their life quality. Clinically, these patients need long-term follow-ups routinely with palpation and ultrasonography, which are subjective and qualitative diagnosis. In this study, we proposed automatic segmentation for muscle/fibers ultrasound images to help doctors objectively and quantitatively evaluate the damage condition of muscle, optimize treatment strategies and rehabilitation courses, and predict prognosis after RT.
Materials and Methods
In this study, 28 patients with head/neck cancer, 20 males and 8 females ranging from 28 to 69 years old, received 66~70 Gy during radiation treatment, and their B-mode ultrasound images were acquired before the treatment and during one year follow-ups. The proposed method was mainly divided into two parts: muscle segmentation and fiber quantification. It was firstly to train a U-Net to segment the sternocleidomastoid muscle (SCM) as ROI input for the next step, and secondly Gabor filter and Conditional adaptive thresholding methods were applied to segment the fibrotic tissue from the predicted SCM ROIs. In addition, we refined the segmentation of fibrotic tissue using One-Class Support vector machine (SVM) accounting for both spatial coherence and appearance. Finally, the ratio of fiber to muscle and its variations were calculated to objectively evaluate the severity of muscle fibrosis.
Results
Our results show that U-Net based SCM segmentation can achieve an Accuracy over 99% with a Dice coefficient of over 96%. The Precision and Recall are above 96%. The fiber analysis indicates that the variations of fiber/muscle ratio increase at 3 and 6 months after RT and decrease at 1 year after RT.
Conclusion
The proposed method is an objective and quantitative to segment SCM of the ultrasound image and analyze the condition of its fiber/muscle ratio. It can be used for forecasting the prognosis of muscle fibrosis and may support doctors’ decision-making in treating this symptom.

並列關鍵字

B-mode ultrasound image ； head/neck cancer ； muscle fibrosis ； U-Net ； Gabor filter ； Conditional adaptive thresholding ； One-Class SVM

參考文獻

1. Dorr W and Hendry JH. (2001) Consequential late effects in normal tissues. Radiotherapy and Oncology, 61(3), 223-231.

2. 劉子維. (2014) 應用於放射線所造成肌肉組織傷害之超音波影像分析方法.

國立臺灣大學醫學工程學研究所學位論文.

3. Zhao Y, Haginoya K, and Sun G, et al. (2003) Platelet‐derived growth factor and its receptors are related to the progression of human muscular dystrophy: an immunohistochemical study. The Journal of Pathology, 201(1), 149-159.

4. Hall EJ and Giaccia AJ. (2012) Radiobiology for the Radiologist 7th. Lippincott Williams Wilkins.