TY - JOUR T1 - An end-to-end automatic lesion detection system in whole-body bone scintigraphy by deep learning. JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 142 LP - 142 VL - 62 IS - supplement 1 AU - Shengyun Huang AU - Kaibin Huang AU - Xue Li AU - Shawn Li AU - Yi Gao AU - Ying Liang Y1 - 2021/05/01 UR - http://jnm.snmjournals.org/content/62/supplement_1/142.abstract N2 - 142Objectives: Due to the limitation of image resolution and visual contrast, bone scintigraphy interpretation is susceptible to subjective factors, such as experience of physicians. This considerably affects the accuracy and repeatability of lesion detection. We design and implement an automatic lesion detection system in whole-body bone scintigraphy by deep learning. Methods: A total of 235 patients with positive lesions in bone scintigraphy are retrospectively analyzed. The whole-body bone scans were obtained 3-4 h after an intravenous injection of 99mTc-methylene diphosphonate (740-1110 MBq) with a γ-camera equipped with low-energy, high-resolution parallel-hole collimators (scan speed, 20 cm/min; matrix, 256×1024). DICOM images with 16bit depth include both anterior and posterior views. All positive lesions are annotated by board certificated physicians. These images were randomly separated into training and testing groups (Training:Testing=194:21 for anterior and 129:16 for the posterior). The backbone of our system is derived from U-net segmentation framework (Figure 1). The raw images are cropped into a patch of 128×128 matrix and fed to the supervised neural network. Weighted binary cross entropy loss is employed to improve the detection sensitivity whereas the dice index is utilized to evaluate the performance of lesions detection. Higher dice index indicates more consistency between the ground truth and the segmentation results. Detection rate and precision are calculated to estimate the global lesion detection performance. Results: Among 235 patients, 365 hotpots of 215 anterior images and 304 hotpots of 145 posterior images are annotated in total (Figure 2). For anterior images, the system achieves dice index, detection rate and precision of 0.73 (0.52-0.90), 88.2% (322/365), and 79.1% (322/407), respectively (Table 1). For the posterior images, the values are, respectively, 0.89 (0.62-0.97), 94.7% (288/304), and 84.0% (288/343) (Table 2). Conclusions: At present, most algorithms only perform local bone scan image analysis, such as in the chest region. We propose an end-to-end deep learning framework that segments all positive lesions in whole-body bone scans. Our results indicate that the proposed system performs well in both anterior and posterior images, with proper detection rate and precision. View this table:Table 1 Lesion detection results for anterior images View this table:Table 2 Lesion detection results for posterior images ER -