一致性結構單元點群漂移：放射線治療肺部電腦斷層掃瞄之多時間點影像對位演算法

Abstract

過去數年來惡性腫瘤一直高居國內十大死因之首位，其中肺癌不分性別高居所有癌症死亡原因第一位，占所有死亡人數的28.3%。早期肺癌病患適合手術切除的僅佔總體約15%，大多數之肺癌被發現時肺腫瘤期別已經進入中晚期，此時放射治療為主要治療方式。在放射線治療過程中，高能放射線除了對腫瘤進行治療，同時也會傷害鄰近的肺部正常組織導致肺部實質病變(radiation induced lung diseases, RILD)。RILD 對病人的傷害有潛在性致命風險，並影響病人治療後生活品質。因此，若能從治療前得知劑量區與RILD的關聯性，進而預測治療後RILD嚴重性，即可協助醫師調整治療計畫避免嚴重的併發症危及病人性命。 因為肺部輻射敏感度(radiosensitivity)的差異，使得每一位病人的腫瘤與肺部不同組織位置對放射治療的反應有所不同。本研究的終極目的是協助臨床醫師觀察不同時間點在相同劑量分佈中RILD變化的關係，於放射線治療中，將個體差異列入考量，以有效降低RILD的損傷。為了達成此一目的，本論文提出一多時間對位技術，容許觀察治療前後相同區域之影像變化，以便探討治療後不同時間點同一區域之RILD差異。 在過去的研究中，本實驗室基於Coherence Point Drift(CPD)發展了非剛性點群對位演算法，透過肺區以外鄰近肺臟的解剖結構，描述不同時間點呼吸飽滿度所造成肺部擴張差異。然而CPD的基本假設是點群位移控制具有同致性(coherence)，但吸氣時肋骨抬升、肺部擴張以及心臟跳動等各自解剖構造形變的模式非同致性(incoherence)，導致靠近心臟以及氣管邊緣的對位錯誤率增加，因此使用CPD用來描述整個肺區周圍解剖特徵形變有其不足之處。 為了提升肺壁內緣吻合度以及更加符合每個解剖構造運動模型的差異，本研究提出一個用於克服RILD病變之Component-Structure Coherence Point Drift(CSCPD)多時間點對位演算法，其特色在於依據使用者之定義將整體構造拆解成多個解剖構造單元，各別進行CPD對位，相較於使用整個肺區結構描述整體運動形變更符合各自解構運動模型。本論文中依據結構組成將分成兩部分探討，第一:採用肺區以外鄰近肺臟的解剖結構作為非剛性對位的形變依據，包含肋骨、脊椎、氣管、肺下緣、肺區內緣、胸骨柄等，其中以脊椎作為剛性對位的參考結構，並將各自單元結構進行CPD點群對位，主要評估肺壁邊緣對位好壞，定義為CSCPD_W (CSCPD_LungWall)。第二: 本研究提出區域血管點群匹配法，是一個以少數血管匹配點能夠產生大範圍足夠的代表性區域血管樹控制點的方法。除了針對肺區外周圍解剖構造以外，同時加入肺區內足夠代表性的對位參考點，定義為CSCPD_WV(CSCPD_LungWall and Vessel)。最後利用TPS將其各自解構單元點群合併後求得分析區域內其餘各點形變後的位置。 本研究透過獨立樣本t檢定分析CPD、CSCPD_W肺區邊緣吻合錯誤值集合的平均值，共18名輕中度嚴重等級的RILD病患，其分析結果CSCPD_W與CPD之間有達到臨床上顯著差異(P<0.05)，說明本研究所提出的CSCPD_W各自結構單元運動模型的對位概念能有效解決CPD對位所面臨非同致性(incoherence)的問題。接著進一步使用重複測量變異數分析(One-way ANOVA with repeated measures)，利用成對比較檢定CPD、CSCPD_W、CSCPD_WV三種對位方法之間治療前後CT影像相對應20個血管匹配點Target Registration Error平均值差異是否顯著，共六名RILD病患，其分析結果CSCPD_WV相對於CSCPD_W以及CSCPD 和CPD兩兩比較皆有達到顯著差異(P<0.01)，說明區域血管點群匹配法能有效提升內部血管對位準確率。 本論文提出透過各自結構單元運動模型之CSCPD多時間點對位技術，有效克服放療後肺實質的改變造成肺區對位的困難。其CSCPD相對於CPD對位方法能有效降低肺壁邊緣錯誤值以及提升肺區內血管對位精準度。

Malignant neoplasms have been the top leading cause of death in Taiwan for decades. Lung cancer, regardless of gender, ranks the first leading cause of cancer deaths, amounting to 28.3% of all cancer deaths. Surgery is the best and effective method in the early stage of the lung cancer. However, only 15% of the diagnosed early-stage patients are suitable for surgery. Most lung cancers, when found, have been in the intermediate to terminal stages, for which radiation therapy usually serves as one of the main therapeutic approaches. During radiation therapy, high intensity radiation not only kills tumor cells but simultaneously causes damage to normal lung tissues resulting in radiation induced lung disease (RILD). RILD is a severe complication of radiotherapy in lung cancer patients, which poses potential threat to life and deteriorates patients’ quality of life. If the correlation between the radiation dose distribution and RILD is known and a prediction model of RILD can be built before radiotherapy, it would be of great help for clinicians to make the therapeutic plan and prevent severe complications. Due to the variation of pulmonary radiosensitivity among different patients, the tumors and lung tissues of different patients may respond to the radiation therapy differently. The ultimate goal of this study was to assist the medical doctors in understanding the correlation between the RILD changes and dose distribution over time such that the individual difference could be taken into account in the effort of minimizing the tissue damages of RILD during radiation therapy. To achieve this goal, a new longitudinal registration algorithm was proposed in this thesis, which allowed a medical doctor to investigate the longitudinal RILD changes of the same regions before and after radiation therapy. To establish the corresponding pixels of the pulmonary CT images before and after radiation therapy, our laboratory had developed a non-rigid longitudinal registration algorithm based on the Coherence Point Drift (CPD) algorithm previously using the structures surrounding the lung as a whole to describe the deformation resulted from breathing. While this algorithm was able to alignment two lung CT images to a reasonable extent, it suffered the problem that the lung deformation does not satisfy the basic assumption of the CPD method, i.e., all points moving coherently as a group. In reality, different portions of the lung structures, including lung walls and parenchyma, have different deformation models, i.e., deforming incoherently. As a result, larger misalignment errors were observed for the portions of the lung wall around the heart, which suggested that it is insufficient to model lung deformation using the CPD method. To improve the registration accuracy of lung boundaries and account for the incoherent motion nature of different portions of the lung structures, a Component-Structure CPD (CSCPD) registration algorithm was proposed in this thesis. The basic idea of the CSCPD algorithm was that the lung was decomposed into several component structures, each of which was then registered by the CPD method. Compared with the CPD algorithm modeling the whole deformation as a group, the CSCPD algorithm could better describe the individualized coherent deformation model of each component structure. To present the ideas and the performances of the CSCPD algorithm, two versions of the CSCPD algorithms were described in this thesis according to the anatomical compositions. The first one, denoted as CSCPD_W (CSCPD_LungWall), focused on the registration of the lung wall. The CSCPD_W algorithm first performed a rigid registration using the spine as the reference structure. Following that, multiple non-rigid registrations were carried out for such structures as ribs, spine, airway, innerr-lung wall, lower-lung wall, and sternum, each of which served as a component structure registered by a CPD method. The second one, denoted as CSCPD_WV(CSCPD_LungWall and Vessel), was an augmented version of the CSCPD_W by integrating a new idea, namely, regional vascular point matching, to account for the spatially-variant deformation within the parenchyma. The uniqueness of the regional vascular point matching method lied in the capability of generating a great number of corresponding point pairs with only tens of pre-selected corresponding point pairs. Finally, the thin-plate spline (TPS) method was used to construct the non-rigid registration model for the whole lung. Independent two-sample t test was used to compare the mean registration errors of lung boundaries achieved by the CPD and CSCPD_W non-rigid registration algorithms for 18 patients with mild-to-severe RILD. The results showed that the CSCPD_W algorithm was significantly better than the CPD method (p<0.05), which suggested that the proposed CSCPD algorithm was effective in resolving the incoherent motion problem encountered by the CPD method. To assess the registration performances of the CPD, CSCPD_W and CSCPD_WV algorithms on the parenchyma, One-way ANOVA with repeated measures was employed to compare the target registration error (TRE) of 20 manually selected blood vessels landmarks on the pre- and post-treatment CT images of 6 patients. While the null hypothesis of the ANOVA was rejected, the pair-wise post-hoc comparison tests showed that the mean TRE of CSCPD_WV was significantly smaller than those of CSCPD_W and CPD, respectively (p<0.01). It implied that the idea of regional vascular point matching can improve the registration accuracy within the parenchyma. In summary, the thesis proposed a new non-rigid registration algorithm, called CSCPD, for the CT images of the same patient before and after radiation therapy. The key idea of the CSCPD algorithm was to decompose the lung, including the lung wall and parenchyma, and the peripheral structures into component structures and register each of them by the CPD method. The corresponding point pairs established by all component structures were then used to construct the non-rigid registration model using TPS. The analysis results confirmed that the CSCPD algorithm significantly enhance the registration accuracy for both of the lung wall and parenchyma.