Go to AFOMP Pulse Contents/Index Page
(Xin Xie, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China)
Background and Objectives: For breast cancer postoperative radiotherapy, the target volume definition of tumor bed is susceptible to the number of surgical clips, clarity of seroma, inter-observer variability and other factors. Registering preoperative and postoperative images and calculating the tumor contour propagation can help target volume definition. Existing researches mainly employed intensity-based registration methods. However, the structures in pre- and post-operative images were non-corresponding because of tumor resection, clip insertion and post-surgical changes. And these non-corresponding structures will affect the results of intensity-based methods. In clinical practice, the target volume of tumor bed is manually delineated by radiation oncologists. However, manual delineation is affected by many factors. It is time-consuming and labor intensive. And there exists obvious inter-observer variability. Besides, tumor bed marker is considered to be a major reference in target volume definition. So far titanium clip has been widely used both at home and abroad. However it was found in clinical applications and research reports that the usage of titanium clip had two main problems. First, it could result in metal artifacts both on simulation CT and MR images. Second, it was non degradable and thus had an effect on postoperative recovery and cosmetic results. To solve these problems, this study is to be carried out in the following three aspects: ⅰ) To establish new image registration method for improving registration accuracy. ⅱ) To establish new auto-segmentation method for improving efficiency and consistency of delineation. ⅲ) To develop biodegradable tumor bed marker for solving the problem of titanium clip.
Materials and Methods: Two deformable image registration (DIR) methods were proposed to register preoperative CT and planning CT images. One method was multi-metric DIR method which combined intensity-based, fiducial-based and region-based metrics together. The other method was two-step DIR method which combined biomechanically based finite element method (FEM) with intensity-based DIR method. This two-step DIR method modeled global organ deformation and local tumor resection induced changes separately.
A deep learning based auto-segmentation method was established to segment target volume of breast tumor bed.The tumor contour in preoperative CT image was first deformed and propogated onto postoperative CT image via image registration. This propogated contour acted as prior information, which provided the initial position of target volume on postoperative CT image. Then a prior informationguided deep learning 3D U-Net was developed.
Biodegradable materials were used to develop tumor bed marker and four steps were successively carried out:ⅰ) The most suitable material was determined by evaluating imaging results on both CT and MR images, as well as changes of appearance, mass and viscosity during in-vitro degradation testing. ⅱ) In consideration of the problem that titanium clip had in clinical use, and the property of selected material itself, the proper fixation mode was determined. Then adapted shape of marker was designed. ⅲ) The ideal size of marker was determined through suture experiment on gauze and siliconeaccording to predefined evaluation criteria. ⅳ) In pre-experiment of animal experiment, the implant model of rat was established first. Then validity assessment based on image scanning experiment and safety assessment based on specimen analysis experiment were performed.
Results: For image registration, the mean TREs of the intensity-based DIR (i-DIR) method for soft tissues, rigid structures and boundaries were 7.82 mm, 3.34 mm, and 6.93 mm. And the multi-metric DIR method (fri-DIR) achieved mean TREs of 2.06 mm, 3.02 mm, and 3.70 mm for corresponding three pointsets. Compared with the i-DIR method, the p-values were less than 0.05 of multi-metric method for soft tissue and boundary alignment, which indicated significance with the introduction of fiducial-based and region-based metrics. Besides, the two-step DIR method achieved mean TREs of 3.01 mm, 3.34 mm, and 6.93 mm for corresponding three pointsets. Compared with the i-DIR method, the p-values were less than 0.05 of the two-step DIR method for soft tissue alignment, which indicated that the improvement achieved by the two-step DIR method was significant. The TRE results in rigid structure and body boundary were the same for the two methods since the two pointsets were outside the corrected region.
For auto-segmentation, the average DSC achieved by prior information guided deep learning network was 0.808.Comparatively, the average DSC achieved by the traditional gray-level thresholding segmentationmethod was 0.622.P value was less than 0.05, which indicated that the improvementachieved bythe proposed auto-segmentation method was significant.
For tumor bed marker, five kinds of biodegradable materials, PolyGlycolic Acid (PGA), Poly-L-Lactic Acid (PLLA), magnesium alloy1(Mg1), magnesium alloy2 (Mg2) and Poly Lactic Acid/Beta-Tricalcium Phosphate (PLA/β-TCP) were initially determined through literature review and manufacturer consulting. Image scanning experiment showed that, test sample made of PLA/β-TCP material did not result in obvious artifacts both on CT and MR images. Also it had distinct difference from surrounding background. And because of the increasement of density, its contrast with surrounding background was notably enhanced on CT images. Besides, in-vitro degradation testing showed that, the appearance of PLA/β-TCP sample did not have substantial change at 40th week.However the viscosity decreased evidently and mass loss rate was about 10%.In terms of degradation, it could remain undegradable within nine months’ observation. Also it entered into an accelerated degradation stage after nine months and would not remain in the body for too long. Considering the degree of imaging distinction and degradation cycle, PLA/β-TCP was determined as the most suitable biodegradable material. The marker was further designed as button-like structure with circular apertures and notch. And the ideal size was decided as 8 mm length, 6 mm width and 2 mm thickness. Subsequent pre-experiment of animal experiment further proved that newly-developed tumor bed marker could meet the imaging requirements. As for safety assessment, some basic data was obtained and the feasibility of formal experiment was demonstrated.
Conclusions: For image registration, compared with the intensity-based DIR method, the two proposed DIR methods both improved registration accuracy. The influence of large deformations and non-correspondence on registration accuracy was reduced to a certain degree. For auto-segmentation, prior information guided deep learning network improved the segmentation accuracy, which could improve efficiency and consistency of delineation. For tumor bed marker, the most suitable biodegradable material, as well as the shape and size of tumor bed marker was determined. Pre-experiment of animal experiment demonstrated the feasibility of formal experiment.