Purpose: Web-based exposure estimation systems are advantageous for estimating exposure doses for computed tomography (CT) scans. However, such systems depend on the imaging conditions of the slices, and a considerable amount of time and effort is needed to select the slices and extract their imaging conditions from the relevant CT volume data. In this study, we used a convolutional neural network (CNN) to automatically classify specific slices from available CT volume data for use by a Web-based exposure estimation system. We also proposed a method to automatically obtain the imaging conditions of these classified slices. The objective of this study was to improve the efficiency of effective dose estimation. Method: We automatically classified specific slices from CT volume data using two different CNN architectures: VGG-16 and Xception. We organized the dataset into 5 categories corresponding to the contents of the specific slices. We also tested a 9-category version in which the slices were supplemented with their adjacent slices. We automatically obtained the imaging conditions from the DICOM tags of the specific slices that were classified from the CT volume data by the CNN and then estimated the effective exposure dose provided by the Web-based exposure estimation system. Result: Using the 5-category dataset approach, the error in the effective exposure dose was 13% for VGG16 and 6% for Xception. When the 9-category approach was used, the error in the effective exposure dose was 0.8% for VGG16 and 0.6% for Xception. In both the architectures, less than 5 minutes was needed in the classification of the specific slices, followed by the extraction of their imaging conditions; however, VGG16 required the shortest processing time. Conclusion: By supplementing a Web-based exposure estimation system with a CNN and adopting our proposed method, we were able to improve the efficiency of effective dose estimation.
Purposes: The purposes of this study were to automatically extract full forms from abbreviations by using Word2vec for terminology expansion and determine the optimal parameters that ensure the highest accuracy. Methods: Approximately 300000 English abstracts on “image diagnosis” were collected using PubMed from January 1994 to December 2018. As preprocessing, all uppercase letters in the collected data were converted to lowercase letters, and symbols were deleted. In addition, compound word recognition was performed using RadLex published by the Radiological Society of North America and the abbreviation collection published by the Japanese Society of Radiological Technology. Next, distributed representations were generated by two algorithms, continuous bag-of-words (CBOW) and Skip-gram, by using the following parameters: iteration numbers (3–85) and dimensions of word vectors (50–1000). Abbreviations were input to the generated distributed representations, and full forms with the highest cosine similarities with the abbreviations were identified. Then, the rates of the correct answers were calculated by comparing the predicted full forms to 214 gold standards extracted from the abbreviation collection. Results: The highest correct answer rate was 74.3% by Skip-gram, 200 dimensions and 10 iterations. This rate was higher in Skip-gram than in CBOW for all the tested conditions. Conclusion: The accuracy of extracting the full forms by Word2vec is 74.3%, and this result contributes to the consistency of a terminology and the efficiency of terminology expansion.
Purpose: Automated analysis of skeletal muscle in whole-body computed tomography (CT) images uses bone information, but bone segmentation including the epiphysis is not achieved. The purpose of this research was the semantic segmentation of eight regions of upper and lower limb bones including the epiphysis in whole-body CT images. Our targets were left and right upper arms, forearms, thighs, and lower legs. Method: We connected two 3D U-Nets in cascade for segmentation of eight upper and lower limb bones in whole-body CT images. The first 3D U-Net was used for skeleton segmentation in whole-body CT images, and the second 3D U-Net was used for eight upper and lower limb bones’ segmentation in skeleton segmentation results. Thirty cases of whole-body CT images were used in the experiment, and the segmentation results were evaluated using Dice coefficient with 3-fold cross-validation. Result: The mean Dice coefficient was 93% in the left and right upper arms, 89% in the left and right forearms, 95% in the left and right thighs, and 94% in the left and right lower legs. Conclusion: Although the accuracy of the segmentation results of relatively small bones remains a challenge, the semantic segmentation of eight regions of upper and lower limb bones including the epiphysis in whole-body CT images has been achieved.
Purpose: The purpose of this study was to propose a method for segmentation and volume measurement of graft liver and spleen of pediatric transplant recipients on digital imaging and communications in medicine (DICOM) -format images using U-Net and three-dimensional (3-D) workstations (3DWS) . Method: For segmentation accuracy assessments, Dice coefficients were calculated for the graft liver and spleen. After verifying that the created DICOM-format images could be imported using the existing 3DWS, accuracy rates between the ground truth and segmentation images were calculated via mask processing. Result: As per the verification results, Dice coefficients for the test data were as follows: graft liver, 0.758 and spleen, 0.577. All created DICOM-format images were importable using the 3DWS, with accuracy rates of 87.10±4.70% and 80.27±11.29% for the graft liver and spleen, respectively. Conclusion: The U-Net could be used for graft liver and spleen segmentations, and volume measurement using 3DWS was simplified by this method.
Purpose: It is well known that there is a trade-off relationship between image noise and exposure dose in X-ray computed tomography (CT) examination. Therefore, CT dose level was evaluated by using the CT image noise property. Although noise power spectrum (NPS) is a common measure for evaluating CT image noise property, it is difficult to evaluate noise performance directly on clinical CT images, because NPS requires CT image samples with uniform exposure area for the evaluation. In this study, various noise levels of CT phantom images were classified for estimating dose levels of CT images using convolutional neural network (CNN). Method: CT image samples of water phantom were obtained with a combination of mAs value (50, 100, 200 mAs) and X-ray tube voltage (80, 100, 120 kV). The CNN was trained and tested for classifying various noise levels of CT image samples by keeping 1) a constant kV and 2) a constant mAs. In addition, CT dose levels (CT dose index: CTDI) for all exposure conditions were estimated by using regression approach of the CNN. Result: Classification accuracies for various noise levels were very high (more than 99.9%). The CNN-estimated dose level of CT images was highly correlated (r=0.998) with the actual CTDI. Conclusion: CT image noise level classification using CNN can be useful for the estimation of CT radiation dose.
Computed tomography (CT) is used for the attenuation correction (AC) of [F-18] fluoro-deoxy-glucose positron emission tomography (PET) image. However, acquisition of a CT image for this purpose requires increasing the radiation dose of the patient. To generate a pseudo-image, a generative adversarial network (GAN) based on deep learning is adopted. The purpose of this study was to generate a pseudo-CT image, using a GAN, for the AC of the PET image, with the aim of reducing the dose of the patient. A set of approximately 15,000 no-AC PET and CT images was used as the training sample, and the CycleGAN was employed as the image generation model. The training samples were inputted in the CycleGAN, and the hyperparameters, i.e., the learning rate, batch size, and number of epochs were set to 0.0001, 1, and 300, respectively. A pseudo-PET image was obtained using a pseudo-CT image, which was used for the AC of the no-AC PET image. The coefficient of similarity between the real and generated pseudo-images was estimated using the peak signal-to-noise ratio (PSNR) , the structural similarity (SSIM), and the dice similarity coefficient (DSC). The average values of PSNR, SSIM, and DSC of the pseudo-CT were 31.0 dB, 0.87, and 0.89, and those of the pseudo-PET were 35.9 dB, 0.90, and 0.95, respectively. The AC for the whole-body PET image could be accomplished using the pseudo-CT image generated via the GAN. The proposed method would be established as the CT-less PET/CT examination.
Purpose: The noise generated in ultra-high-resolution computed tomography (U-HRCT) images affects the quantitative analysis of emphysema. In this study, we compared the physical properties of reconstructed images for hybrid iterative reconstruction (HIR) and deep learning reconstruction (DLR), which are reconstruction methods for reducing image noise. Using clinical evaluation, we evaluated the correlation between low attenuation volume (LAV) % obtained by CT and forced expiratory volume in 1 s per forced vital capacity (FEV1/FVC) obtained by respiratory function tests. Materials and methods: CT data obtained by HIR and DLR were used for analysis (matrix size: 1024´1024, slice thickness: 0.25 mm). The physical characteristics were evaluated for the modulation transfer function (MTF) and noise power spectrum (NPS). Display-field of view (D-FOV) was analyzed by varying between 300 mm and 400 mm. The clinical data evaluated the relationship between LAV% and FEV1/FVC by Spearman’s correlation coefficient. Result: The 10% MTFs were 1.3 cycles/mm (HIR) and 1.3 cycles/mm (DLR) at D-FOV 300 mm, and 1.2 cycles/mm (HIR) and 1.1 cycles/mm (DLR) at D-FOV 400 mm. The NPS had less noise in DLR than HIR in all frequency ranges. The correlation coefficients between LAV% and FEV1/FVC were 0.64 and 0.71, respectively, in HIR and DLR. Conclusion: There was no difference in the resolution characteristics of HIR and DLR. DLR had better noise characteristics than HIR. The correlation between LAV% measured by HIR and DLR and FEV1/FVC is equivalent. The noise characteristics of the DLR enable the reduction of exposure to emphysema quantitative analysis by CT.
Purpose: Volumetric modulated arc therapy (VMAT) can acquire projection images during rotational irradiation, and cone-beam computed tomography (CBCT) images during VMAT delivery can be reconstructed. The poor quality of CBCT images prevents accurate recognition of organ position during the treatment. The purpose of this study was to improve the image quality of CBCT during the treatment by cycle generative adversarial network (CycleGAN). Method: Twenty patients with clinically localized prostate cancer were treated with VMAT, and projection images for intra-treatment CBCT (iCBCT) were acquired. Synthesis of PCT (SynPCT) with improved image quality by CycleGAN requires only unpaired and unaligned iCBCT and planning CT (PCT) images for training. We performed visual and quantitative evaluation to compare iCBCT, SynPCT and PCT deformable image registration (DIR) to confirm the clinical usefulness. Result: We demonstrated suitable CycleGAN networks and hyperparameters for SynPCT. The image quality of SynPCT improved visually and quantitatively while preserving anatomical structures of the original iCBCT. The undesirable deformation of PCT was reduced when SynPCT was used as its reference instead of iCBCT. Conclusion: We have performed image synthesis with preservation of organ position by CycleGAN for iCBCT and confirmed the clinical usefulness.