Receive a weekly summary and discussion of the top papers of the week by leading researchers in the field.

In Frontiers in medicine

Deep learning technology is now used for medical imaging. YOLOv2 is an object detection model using deep learning. Here, we applied YOLOv2 to FDG-PET images to detect the physiological uptake on the images. We also investigated the detection precision of abnormal uptake by a combined technique with YOLOv2. Using 3,500 maximum intensity projection (MIP) images of 500 cases of whole-body FDG-PET examinations, we manually drew rectangular regions of interest with the size of each physiological uptake to create a dataset. Using YOLOv2, we performed image training as transfer learning by initial weight. We evaluated YOLOv2's physiological uptake detection by determining the intersection over union (IoU), average precision (AP), mean average precision (mAP), and frames per second (FPS). We also developed a combination method for detecting abnormal uptake by subtracting the YOLOv2-detected physiological uptake. We calculated the coverage rate, false-positive rate, and false-negative rate by comparing the combination method-generated color map with the abnormal findings identified by experienced radiologists. The APs for physiological uptakes were: brain, 0.993; liver, 0.913; and bladder, 0.879. The mAP was 0.831 for all classes with the IoU threshold value 0.5. Each subset's average FPS was 31.60 ± 4.66. The combination method's coverage rate, false-positive rate, and false-negative rate for detecting abnormal uptake were 0.9205 ± 0.0312, 0.3704 ± 0.0213, and 0.1000 ± 0.0774, respectively. The physiological uptake of FDG-PET on MIP images was quickly and precisely detected using YOLOv2. The combination method, which can be utilized the characteristics of the detector by YOLOv2, detected the radiologist-identified abnormalities with a high coverage rate. The detectability and fast response would thus be useful as a diagnostic tool.

Kawakami Masashi, Hirata Kenji, Furuya Sho, Kobayashi Kentaro, Sugimori Hiroyuki, Magota Keiichi, Katoh Chietsugu


YOLOv2, computer vision, deep learning, object detection, positron emission tomography