Accurate image interpretation of Waters’ and Caldwell view radiographs used for sinusitis screening is challenging. Therefore, we developed a deep learning algorithm for diagnosing frontal, ethmoid, and maxillary sinusitis on both Waters’ and Caldwell views. The datasets were selected for the training and validation set (n = 1403, sinusitis% = 34.3%) and the test set (n = 132, sinusitis% = 29.5%) by temporal separation. The algorithm can simultaneously detect and classify each paranasal sinus using both Waters’ and Caldwell views without manual cropping. Single- and multi-view models were compared. Our proposed algorithm satisfactorily diagnosed frontal, ethmoid, and maxillary sinusitis on both Waters’ and Caldwell views (area under the curve (AUC), 0.71 (95% confidence interval, 0.62–0.80), 0.78 (0.72–0.85), and 0.88 (0.84–0.92), respectively). The one-sided DeLong’s test was used to compare the AUCs, and the Obuchowski–Rockette model was used to pool the AUCs of the radiologists. The algorithm yielded a higher AUC than radiologists for ethmoid and maxillary sinusitis (p = 0.012 and 0.013, respectively). The multi-view model also exhibited a higher AUC than the single Waters’ view model for maxillary sinusitis (p = 0.038). Therefore, our algorithm showed diagnostic performances comparable to radiologists and enhanced the value of radiography as a first-line imaging modality in assessing multiple sinusitis.

Blood vessel segmentation (BVS) of 3D medical imaging such as computed tomography and magnetic resonance angiography (MRA) is an essential task in the clinical field. Automation of 3D BVS using deep supervised learning is being researched, and U-Net-based approaches, which are considered as standard for medical image segmentation, are proposed a lot. However, the inherent characteristics of blood vessels, e.g., they are complex and narrow, as well as the resolution and sensitivity of the imaging modalities increases the difficulty of 3D BVS. We propose a novel U-Net-based model named Spider U-Net for 3D BVS that considers the connectivity of the blood vessels between the axial slices. To achieve this, long short-term memory (LSTM), which can capture the context of the consecutive data, is inserted into the baseline model. We also propose a data feeding strategy that augments data and makes Spider U-Net stable. Spider U-Net outperformed 2D U-Net, 3D U-Net, and the fully convolutional network-recurrent neural network (FCN-RNN) in dice coefficient score (DSC) by 0.048, 0.077, and 0.041, respectively, for our in-house brain MRA dataset and also achieved the highest DSC for two public datasets. The results imply that considering inter-slice connectivity with LSTM improves model performance in the 3D BVS task.

The performance of deep learning algorithm (DLA) to that of radiologists was compared in detecting low contrast objects in CT phantom images under various imaging conditions. For training, 10,000 images were created using American College of Radiology CT phantom as the background. In half of the images, objects of 3–20 mm size and 5–30 HU contrast difference were generated in random locations. Binary responses were used as the ground truth. For testing, 640 images of Catphan^® phantom were used, half of which had objects of either 5 or 9 mm size with 10 HU contrast difference. Twelve radiologists evaluated the presence of objects on a five-point scale. The performances of the DLA and radiologists were compared across different imaging conditions in terms of area under receiver operating characteristics curve (AUC). Multi-reader multi-case AUC and Hanley and McNeil tests were used. We performed post-hoc analysis using bootstrapping and verified that the DLA is less affected by the changing imaging conditions. The AUC of DLA was consistently higher than those of the radiologists across different imaging conditions (p < 0.0001), and it was less affected by varying imaging conditions. The DLA outperformed the radiologists and showed more robust performance under varying imaging conditions.