TY - JOUR T1 - <strong>Unsupervised lesion detection in bone scintigraphy using deep learning-based image inpainting technology</strong> JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 403 LP - 403 VL - 60 IS - supplement 1 AU - Seung Kwan Kang AU - Hongyoon Choi AU - Sohyun Park AU - Seok-ki Kim AU - Tae-sung Kim AU - Jae Sung Lee Y1 - 2019/05/01 UR - http://jnm.snmjournals.org/content/60/supplement_1/403.abstract N2 - 403Objectives: Although deep-learning-based abnormal lesion detection algorithms for nuclear medical images are being actively investigated, conventional methods require time- and labor-intensive lesion labeling for supervised network training. This study develops a new method for unsupervised lesion detection in planar bone scintigraphy using deep-learning-based image inpainting technology without requiring manual lesion labeling for deep network training. Methods: Whole-body bone scintigraphic images (anterior/posterior pairs with matrix dimensions of 1024 x 256 x 2) of 10,000 patients were retrospectively analyzed. These were randomly divided into 9,000 and 1,000 cases for training and testing, respectively. The input to the deep neural network for image inpainting was a randomly cropped 256 x 256 x 2 matrix and included a randomly placed 25 x 25 x 2 zero mask (i.e., the unknown region that the network must recover). The network was trained to inpaint (fill) the masked region by using the original 256 x 256 x 2 matrix without applying the mask as a reference. Because most of the data for network training were selected from regions with normal physiological uptake images and normal uptake distributions, the inpainted images tended to have normal uptake patterns. Therefore, by subtracting the inpainted image from the original, we could identify the abnormal uptake pattern. The deep network developed in this study had a U-net-like structure with dense-net sub-blocks and shortcut connections. To promote flexible network training, we introduced novel multiplicative terms to the shortcut connections. To train the network, we minimized L1 loss between the network output and reference as well as perceptual loss using a pre-trained VGG19 network. For the evaluation, the trained network was applied to the whole-body images using a moving mask with stride 7. Results: Following training, the proposed network successfully filled the zero masks in the input images. Moreover, the abnormal uptake pattern in the input data was effectively eliminated using the inpainting technique and was easily detected in subtraction images. Conclusions: The proposed inpainting network generated normal-like images by removing abnormal uptake patterns in the input data. This enabled abnormal lesion detection in whole-body bone scans without manual lesion labeling. ER -