RT Journal Article
SR Electronic
T1 Fully automatic extraction and normalization of brain volume from oncologic FDG PET scan using deep learning
JF Journal of Nuclear Medicine
JO J Nucl Med
FD Society of Nuclear Medicine
SP 1423
OP 1423
VO 61
IS supplement 1
A1 Wonseok Whi
A1 Hongyoon Choi
A1 Jin Chul Paeng
A1 Gi Jeong Cheon
A1 Keon Wook Kang
A1 Dong Soo Lee
YR 2020
UL http://jnm.snmjournals.org/content/61/supplement_1/1423.abstract
AB 1423Introduction: FDG PET has been playing a crucial role in the diagnosis of several types of cancer. While the brain volume is often covered in the oncologic PET scan, the analysis of the brain typically depends on visual inspection. This routine rough visual inspection has limited the utilization of quantitative information of brain PET, which can be associated with incidental brain disorders as well as metastasis. With the aid of automated analytic methods such as statistical parametric mapping, it would be possible to get additional quantitative information on brain metabolism. Here, we developed a deep learning-based fully-automated brain extraction method from oncologic PET studies. Methods: As a proof-of-concept study, 500 FDG-PET scans in the oncology clinic were retrospectively collected. The brain extraction model had two objectives: 1) the evaluation of whether a scan included the entire brain and 2) 3-dimensional bounding box which included brain from a PET scan. To achieve them simultaneously, maximum intensity projection (MIP) images of each 500 PET scans were firstly generated. Two-dimensional bounding boxes were manually drawn on the anterior and lateral views of MIP images, respectively. By merging coordinate data from bounding boxes drawn on the two MIP images, coordinates of 3-D bounding boxes were obtained. A deep learning model based on convolutional neural networks was developed to predict 3-D coordinates of bounding boxes using two MIP images. We validated the model by randomly selected 10% of the PET data. From the PET data and predicted coordinates of bounding boxes, we extracted brain images from whole body PET and spatially normalized to the template space. Results: The deep learning-based brain extractor successfully identified the existence of whole-brain volume. The accuracy for predicting the existence of whole-brain volume in the validation set was 98%. The performance of extracting the brain measured by the intersection-over-union (IOU) of 3-D bounding boxes was 72.9±12.5% for the validation set. Using coordinates of predicted bounding boxes, all brains were successfully cropped and automatically normalized into the template space. Conclusions: As a proof-of-concept study, we show that quantitative assessment of brain images is feasible by automatically extracting brain from oncologic FDG PET studies using a deep learning model. The automatically analyzed quantitative information on brain metabolism may be utilized in various purposes to identify abnormal findings related to brain disorders as well as incidental metastasis.