Abstract
3060
Introduction: Bone metastases are widely found in many malignant tumors, and an accurate and efficient delineation approach is desired in the routine clinical diagnosis. In this study, a direct and effective method for automatic bone metastases delineation was developed and tested.
Methods: This study included a dataset of 5 patients (mean age 64.2 years, range 55-71 years) with metastatic castration-resistant prostate cancer (mCRPC), who were examined at Renji Hospital, Shanghai Jiao Tong University School of Medicine. Sixty-minute 68Ga-PSMA dynamic PET scan was performed for each patient on the uEXPLORER scanner (United imaging healthcare, Shanghai, China). A low-dose CT ( 1mSv) was obtained for attenuation correction. PET raw data were acquired in 3D list mode. The ordered subset expectation maximization (OSEM) (4 iterations 20 subsets) with point spread function (PSF) and time of flight (TOF) algorithm was used (360x360, slice thickness 2.89 mm, Field of view (FOV) 600 mm) for PET imaging reconstruction. The automatic delineation method includes the following steps: Firstly, CT image segmented into regions over/below the threshold of 116 based on Hounsfield Unit (HU), which will be marked as ‘1’ or ‘0’. Similarly, the lesions in PET image are segmented by a threshold of 40% of max standard uptake value (SUVmax), which is marked as ‘1’ or ‘0’ as the region value over or below the threshold. The bone metastasis regions were then obtained by multiplying the generated binary mask and segmented bone mask. Secondly, the PET/CT image is fused by PET mapping image with CT map image, Thus only the double‘1’region will be left as the non-zero area, which will be taken as the masked bone metastases in the fused PET/CT image. The algorithm flowchart in detail is illustrated in Figure 1. To test the proposed algorithm, 5 prostate cancer patients with bone metastases who underwent 68Ga-PSMA PET/CT were evaluated in this study.
Results: The tumors delineated by the proposed algorithm is well matched with the manual clinical standard SUV tumor delineation as showed in the Figure 2. The bone metastases are accuracy delineated by the automatic segmentation algorithm. Meanwhile, the images defined by the proposed algorithm only showed the bone and tumor lesion on the bone without background interference, which was superior to the standard SUV images. For all the labeled lesions, the algorithm can achieve 99% precision,99% recall, and F1 score of 99%.
Conclusions: In conclusion, the bone tumors were obviously visualized in the fused PET/CT. And this background is clear with pretty low noise and only tumor lesion and bone were masked. This automatic segmentation approach is valuable application in assisting clinical routine diagnosis in bone tumor and bone metastases.