利用三維形態分析診斷肺臟腫瘤之系統

The 3D morphological analysis and diagnosis system for pulmonary nodule

葉嘉芬(Chia-Fen Yeh)

指導教授 ： 蘇振隆

中原大學/工學院/醫學工程研究所/碩士(2003年)

https://doi.org/10.6840/cycu200300137

摘要

肺癌為近年來全球性與日遽增的致命疾病之一。由於肺臟的病變相當複雜，因此，利用高解析度的胸腔電腦斷層(Computed tomography, CT)影像可協助胸部X光片上無法顯現的病灶。本研究計畫發展一套利用電腦輔助量化分析及診斷系統，協助醫師鑑別肺部腫瘤，並能提供病人治療前後之病情追蹤。
本研究直接擷取肺部DICOM影像來進行分析，針對所欲觀察之組織部位運用窗階技術(window/level)做調節。透過影像處理技術，如：影像增強、半自動邊緣選取、高斯平滑、曲率計算、似圓性和紋理分析等函數來分析及診斷肺臟腫瘤之形態。此外，亦配合OpenGL和Marching cube演算法表現出胸部與腫瘤Surface render三維影像，提供形態學之觀察。研究過程中，利用影像模擬腫瘤假體，包括不同切片厚度之球形假體CT序列影像與不同外觀之球形假體，來測試系統之參數準確性，並針對臨床肺癌病例進行實際診斷鑑別及探討。此外，也藉由臨床醫師之協助比較系統與傳統人工鑑別腫瘤之良惡性的差異。
初步結果顯示，在假體測試評估中證實系統之準確性，針對不同切片厚度所造成的曲率計算之誤差，乃是因為重建時之插值技術所導致。在實際臨床病例分析診斷時，亦分別針對兩種不同切片厚度3mm和5mm臨床病例進行評估。其結果顯示：切片厚度為3mm之肺腫瘤測試病例中，鑑別正確率為0.67，敏感度為1和系統之信賴度Kappa值為0.378，而導致正確率與Kappa值降低之原因為測試病例數過少，且測試病例中之良性病例大部分皆為發炎型及纖維化之腫瘤，其各項特徵與惡性腫瘤極為相似，因此導致本系統診斷錯誤，而降低正確率；切片厚度為5mm之肺腫瘤測試病例中，鑑別正確率為0.833，敏感度為1和系統之信賴度Kappa值為0.515。研究發現，良性腫瘤之鈣化結節、細菌性感染與纖維化之類型，與惡性腫瘤之特徵類似，因此導致系統之正確率降低。
整體而言，系統之完成能夠提供肺臟腫瘤之形態分析與診斷，並能實際輔助臨床上之肺腫瘤鑑別診斷，亦可提供肺癌治療前後之評估。未來可增加臨床病例之收集與增加其他腫瘤特徵參數，使倒傳遞類神經網路之訓練更加完善以提高系統之正確率，亦可加入自動邊緣偵測來提高系統之效率。

關鍵字

胸腔電腦斷層 ； DICOM ； 曲率計算 ； 倒傳遞類神經網路 ； 紋理分析

並列摘要

Lung cancer becomes one of worldwide deadly diseases growing with each passing day. Using high resolution Computed Tomography images due to complex lung pathological changes can help to find lurk nidus on chest X-ray images. This study developed a set of utilizing computer assistant quantification analysis and diagnosis system in order to help doctors distinguish lung nodules and provide patients with condition track.
This study directly caught lung DICOM images to analyze which where aimed at tissues interesting applying window/level to modulate. Using image process technology, for example, image enhance, semi-automated edge selection, Gaussian smoothing, curvatures calculation, circularity and texture count, etc. the shapes of lung nodules were analyzed and diagnosed. Furthermore, combining OpenGL and Marching cube algorithms could show 3-dimension images of chest and nodules and provide the observation in morphology. During study processing, image simulation nodule phantoms including different sections of thickness spherical phantom CT images and different appearance spherical phantoms were used to test parameters accuracy of system and precede actual diagnostic differentiation and conference for clinical lung cancer cases. Besides, with assistance of clinical doctors, the differences between this system and traditional artificial distinguishing the benignancy and malignant nodules were compared.
Preliminary results showed that the accuracy of system and the error of curvatures calculation in different sections of thickness in phantom test evaluating which due to insert-value technology in reconstruction. In clinical cases analysis, two different sections of thickness 3mm and 5mm were evaluated. The results showed that the accuracy was 0.67, sensitivity was 1 and Kappa is 0.378 for 3mm section thickness lung nodules testing cases. The reasons of low accuracy and Kappa were the testing data size were too small and the great part in benign cases which were inflammation type and fibered nodules, its each characteristics, were similar to malignant nodules. For 5mm section thickness lung nodules testing cases, accuracy was 0.833, specificity was 1 and Kappa was 0.515. For the results, we found that the characteristics of the calcification nodes of benign nodules, bacterial affection and fibered types were similar to malignant nodules, hence, this lower accuracy were got in this system.
The accomplishment of system can provide texture analysis and diagnoses in lung nodules, and help differential diagnosis in clinical lung nodules actually. It also supported the evaluation in lung cancer before and after of treatment. In the future, we will increase the clinical cases collections and feature parameters of other nodules which will perfect the trainings in Back Propagation Neural Network and raise the accuracy of system. To add automatic edge detection function also can improve the efficiency of this system.

並列關鍵字

Curvature calculated ； Computed tomography ； Texture analysis ； Back Propagation Neural Network ； DICOM

參考文獻

[13]趙一平，“肺臟斷層影像之電腦輔助分析”，中原大學醫學工程研究所碩士論文，民國九十一年。

[17]陳逸雯，“肺部腫瘤偵測之電腦輔助診斷系統”，中原大學醫學工程研究所碩士論文，民國九十一年。

[5] M. D. Seemann, A. Staebler, T. Beinert, H. Dienemann, B. Obst, M. Matzko, C. Pistitsch, M. F. Reiser, “ Usefulness of morphological characteristics for the differentiation of benign from malignant solitary pulmonary lesions using HRCT” , European Radiology 1999, vol. 9 , issue 3 , p409~417.

[6] T. Tozaki, Y. Kawata, N.Niki, et al., “Pulmonary organs analysis for differential diagnosis based on thoracic thin-section CT images”, IEEE Transaction on nulclear science, December 1998, p3075~3082.

[7]Y. Kawata, N. Niki, H. Ohmatsu, R.Kakinuma, K. Eguchi, M. Kaneko, N. Moriyama, “Quantitative surface characterization pulmonary nodules based on thin-section CT image”, IEEE Trans. Nuclear Science 1998, vol.45, p2132~2138.