辻本 正和

OBJECTIVE: This study proposes an automated classification of benign and malignant in highly integrated regions in bone single-photon emission computed tomography/computed tomography (SPECT/CT) using a three-dimensional deep convolutional neural network (3D-DCNN). METHODS: We examined 100 regions of 35 patients with bone SPECT/CT classified as benign and malignant by other examinations and follow-ups. First, SPECT and CT images were extracted at the same coordinates in a cube, with a long side two times the diameter of a high concentration in SPECT images. Next, we inputted the extracted image to DCNN and obtained the probability of benignity and malignancy. Integrating the output from DCNN of each SPECT and CT image provided the overall result. To validate the efficacy of the proposed method, the malignancy of all images was assessed using the leave-one-out cross-validation method; besides, the overall classification accuracy was evaluated. Furthermore, we compared the analysis results of SPECT/CT, SPECT alone, CT alone, and whole-body planar scintigraphy in the highly integrated region of the same site. RESULTS: The extracted volume of interest was 50 benign and malignant regions, respectively. The overall classification accuracy of SPECT alone and CT alone was 73% and 68%, respectively, while that of the whole-body planar analysis at the same site was 74%. When SPECT/CT images were used, the overall classification accuracy was the highest (80%), while the classification accuracy of malignant and benign was 82 and 78%, respectively. CONCLUSIONS: This study suggests that DCNN could be used for the direct classification of benign and malignant regions without extracting the features of SPECT/CT accumulation patterns.

OBJECTIVE: Quantitative analyses of gamma-ray accumulation in single-photon emission computed tomography (SPECT), and the evaluation of antiresorptive agent-related osteonecrosis of the jaw (ARONJ) have been reported recently. However, the relationship between the quantitative parameters calculated from SPECT and the detailed morphological changes observed in computed tomography (CT) remains unclear. This study aimed to investigate patients' characteristics and morphological changes observed on CT, and their effects on the quantitative values in SPECT. METHODS: From April 2017 to March 2019, patients diagnosed with ARONJ at our hospital were enrolled. The data obtained before September 2017 were reviewed retrospectively, and other data were collected prospectively. CT scans were evaluated for internal texture, sequestrum formation, periosteal reaction, cortical perforation, bone expansion, and pathological fracture. For quantitative assessment, the ratio of the maximum standardized uptake value (SUV) to the mean SUV in the temporal bone (rSUVmax) was calculated from SPECT images. The factors affecting rSUVmax were investigated by multiple regression analysis. The statistical significance level was set at α = 0.05. RESULTS: Overall, 55 lesions of 42 patients (median age and interquartile range, 75 [67-80 years], 27 female) were evaluated. Male sex (p = 0.007) and bilateral location (p < 0.0001) were selected as variables in the multivariate analysis. Adjusted coefficient of determination R2 was 0.59 (p < 0.0001). CONCLUSION: Sex and horizontal progression of the disease may affect individually calibrated SUVs in SPECT for patients with ARONJ.

The aim of this study was to investigate the relationship of quantitative parameters between the two-dimensional region of interest (ROI) and the three-dimensional volume of interest (VOI) for accumulation of radiopharmaceutical. Single-photon emission computed tomography combined with computed tomography (SPECT/CT) images of the NEMA/IEC phantom were acquired. The ROIs and VOIs were automatically set to the sphere and background in the phantom. We defined as two-dimensional analysis (2D analysis) that which used ROIs set on the center section of the sphere, and as three-dimensional analysis (3D analysis) that which used VOIs set on the center of gravity of the sphere. Dose linearity (DL), the recovery coefficient (RC), the contrast-to-noise ratio (CNR), and standardized uptake value (SUV) were evaluated. Each index value was compared between both analyses. DL was almost 1 under both conditions. RC showed a similar tendency with 2D and 3D analyses. The CNR for 3D analysis was smaller than for 2D analysis. The maximum SUV was almost equal with both analyses. The mean SUV with 3D analysis was underestimated by 4.83% on average compared with 2D analysis. For the same accumulation, a difference may occur in the quantitative index between 2 and 3D analyses. In particular, the quantitative parameters based on the average value tends to be smaller with 3D analysis than 2D analysis. The quantitative parameters in 2D analysis showed dependence upon the cross section used for setting the ROI, whereas 3D analysis showed less dependence on the position of the VOI.

PURPOSE: In recent years, convolutional neural network (CNN), an artificial intelligence technology with superior image recognition, has become increasingly popular and frequently used for classification tasks in medical imaging. However, the amount of labelled data available for classifying medical images is often significantly less than that of natural images, and the handling of rare diseases is often challenging. To overcome these problems, data augmentation has been performed using generative adversarial networks (GANs). However, conventional GAN cannot effectively handle the various shapes of tumours because it randomly generates images. In this study, we introduced semi-conditional InfoGAN, which enables some labels to be added to InfoGAN, for the generation of shape-controlled tumour images. InfoGAN is a derived model of GAN, and it can represent object features in images without any label. METHODS: Chest computed tomography images of 66 patients diagnosed with three histological types of lung cancer (adenocarcinoma, squamous cell carcinoma, and small cell lung cancer) were used for analysis. To investigate the applicability of the generated images, we classified the histological types of lung cancer using a CNN that was pre-trained with the generated images. RESULTS: As a result of the training, InfoGAN was possible to generate images that controlled the diameters of each lesion and the presence or absence of the chest wall. The classification accuracy of the pre-trained CNN was 57.7%, which was higher than that of the CNN trained only with real images (34.2%), thereby suggesting the potential of image generation. CONCLUSION: The applicability of semi-conditional InfoGAN for feature learning and representation in medical images was demonstrated in this study. InfoGAN can perform constant feature learning and generate images with a variety of shapes using a small dataset.

OBJECTIVE: This study aimed to use quantitative values, calculated from bone single photon emission computed tomography (SPECT) imaging, to estimate the reliability of progression evaluation for anti-resorptive agent-related osteonecrosis of the jaw (ARONJ). METHODS: The study population consisted of 21 patients (23 lesions), clinically diagnosed with mandibular ARONJ, who underwent SPECT/CT scanning. Diagnosis and staging of ARONJ were performed according to the American Association of Oral and Maxillofacial Surgeons (AAOMS) definition and the recommendations of the International Task Force on ONJ. Hybrid SPECT/CT imaging quantitative analyses were performed on a workstation. Each volume of interest (VOI) was semi-automatically placed over a lesion with areas of high tracer accumulation, using the GI-BONE® software default threshold method settings. Additionally, control VOI was manually set over an unaffected area. Measured parameters included standardized uptake values (SUV)-maximum (SUVmax) and mean (SUVmean), metabolic bone volume (MBV)-the total volume above the threshold, and total bone uptake (TBU) as calculated by MBV × SUVmean. We also calculated the SUV ratio (rSUV) between the lesion and control area, factoring for differences in individual bone metabolism; the ratios were termed rSUVmax and rSUVmean, accordingly. The product of multiplying the rSUVmean by MBV of a lesion was defined as the ratio of TBU (rTBU). Quantitative values were compared between clinical stages by the Kruskal-Wallis test and subsequent post hoc analysis. RESULTS: MBVs (cm3) were: median, [IQR] Stage 1, 8.28 [5.62-9.49]; Stage 2, 15.28 [10.64-24.78]; and Stage 3, 34.61 [29.50-40.78]. MBV tended to increase with stage increase. Furthermore, only MBV showed a significant difference between clinical stages (p < 0.01). Subsequent post hoc analysis showed no significant difference between stages 1 and 2 (p = 0.12) but a significant difference between stages 2 and 3 (p = 0.048). rSUVmax and rTBU tended to increase with stage increase, but the differences between the stages were not significant (p = 0.10 and p = 0.055, respectively). CONCLUSION: MBV, which includes the concept of volume, showed significant differences between clinical stages and tended to increase with the stage increase. As an objective and reliable indicator, MBV might be an adjunct diagnostic method for staging ARONJ.

It is often difficult to distinguish between benign and malignant pulmonary nodules using only image diagnosis. A biopsy is performed when malignancy is suspected based on CT examination. However, biopsies are highly invasive, and patients with benign nodules may undergo unnecessary procedures. In this study, we performed automated classification of pulmonary nodules using a three-dimensional convolutional neural network (3DCNN). In addition, to increase the number of training data, we utilized generative adversarial networks (GANs), a deep learning technique used as a data augmentation method. In this approach, three-dimensional regions of different sizes centered on pulmonary nodules were extracted from CT images, and a large number of pseudo-pulmonary nodules were synthesized using 3DGAN. The 3DCNN has a multi-scale structure in which multiple nodules in each region are inputted and integrated into the final layer. During the training of multi-scale 3DCNN, pre-training was first performed using 3DGAN-synthesized nodules, and the pulmonary nodules were then comprehensively classified by fine-tuning the pre-trained model using real nodules. Using an evaluation process that involved 60 confirmed cases of pathological diagnosis based on biopsies, the sensitivity was determined to be 90.9% and specificity was 74.1%. The classification accuracy was improved compared to the case of training with only real nodules without pre-training. The 2DCNN results of our previous study were slightly better than the 3DCNN results. However, it was shown that even though 3DCNN is difficult to train with limited data such as in the case of medical images, classification accuracy can be improved by GAN.

PURPOSE: Early detection and treatment of lung cancer holds great importance. However, pulmonary-nodule classification using CT images alone is difficult to realize. To address this concern, a method for pulmonary-nodule classification based on a deep convolutional neural network (DCNN) and generative adversarial networks (GAN) has previously been proposed by the authors. In that method, the said classification was performed exclusively using axial cross sections of pulmonary nodules. During actual medical-examination procedures, however, a comprehensive judgment can only be made via observation of various pulmonary-nodule cross sections. In the present study, a comprehensive analysis was performed by extending the application of the previously proposed DCNN- and GAN-based automatic classification method to multiple cross sections of pulmonary nodules. METHODS: Using the proposed method, CT images of 60 cases with confirmed pathological diagnosis by biopsy are analyzed. Firstly, multiplanar images of the pulmonary nodule are generated. Classification training was performed for three DCNNs. A certain pretraining was initially performed using GAN-generated nodule images. This was followed by fine-tuning of each pretrained DCNN using original nodule images provided as input. RESULTS: As a result of the evaluation, the specificity was 77.8% and the sensitivity was 93.9%. Additionally, the specificity was observed to have improved by 11.1% without any reduction in the sensitivity, compared to our previous report. CONCLUSION: This study reports development of a comprehensive analysis method to classify pulmonary nodules at multiple sections using GAN and DCNN. The effectiveness of the proposed discrimination method based on use of multiplanar images has been demonstrated to be improved compared to that realized in a previous study reported by the authors. In addition, the possibility of enhancing classification accuracy via application of GAN-generated images, instead of data augmentation, for pretraining even for medical datasets that contain relatively few images has also been demonstrated.

OBJECTIVE: Precise prediction of postoperative pulmonary function is extremely important for accurately evaluating the risk of perioperative morbidity and mortality after major surgery for lung cancer. This study aimed to compare the accuracy of a single-photon emission computed tomography/computed tomography (SPECT/CT) method that we recently developed for predicting postoperative pulmonary function versus the accuracy of both the conventional simplified calculating (SC) method and the method using planar images of lung perfusion scintigraphy. METHODS: The relationship between the postoperative observed % values of the forced expiratory volume in 1 second (FEV1) or diffusing capacity for carbon monoxide (DLCO or DLCO') and the % predicted postoperative (%ppo) values of FEV1, DLCO, or DLCO' calculated by the three methods were analyzed in 30 consecutive patients with lung cancer undergoing lobectomy. RESULTS: The relationship between the postoperative observed % values and %ppo values calculated by the three methods exhibited a strong correlation (Pearson r>0.8, two-tailed p<0.0001). The limits of agreement between the postoperative % values and %ppo values did not differ among the three methods. The absolute values of the differences between the postoperative % values and %ppo values for FEV1 and DLCO' were comparable among the three methods, whereas those for DLCO of SPECT/CT were significantly higher than those of the planar method. Conversely, in patients with preoperative %DLCO' of <80% predicted, the absolute values of the differences between the postoperative %DLCO' and %ppoDLCO' of SPECT/CT tended to be smaller than those of the SC and planar methods. CONCLUSION: The accuracy of SPECT/CT for predicting postoperative pulmonary function is comparable with that of conventional methods in most cases, other than in some patients with diffusion impairment.

BACKGROUND: Single-photon emission computed tomography is a tomographic imaging method that acquires a projection image by rotating a gamma camera around by 380° or 180°. For myocardial single-photon emission computed tomography, 180° acquisition is common, but it has limitations including an incomplete reconstruction, which can distort the resulting image. It is possible to produce a complete reconstruction using 360° acquisition, but the testing time is long and is burdensome to patients. METHODS: The nonuniform sampling pitch acquisition (NUSPA) method devised in this study involves reducing the total sampling count using NUSPA that reduces the sampling pitch in the range in which the gamma cameras are closer to the myocardium (RAO45-LPO45) and increases it elsewhere. RESULTS AND CONCLUSION: The NUSPA-1 method based on a 6° sampling pitch had 20 views fewer than 360° acquisition. In addition, the NUSPA-2 method based on a 3.75° sampling pitch had 60 views fewer than 360° acquisition, considerably reducing the testing time. The acquired sinograms from the NUSPA methods were subjected to nonuniform rational B-spline surface interpolation processing, producing data with a uniform sampling pitch, after which image reconstruction was performed. The images after nonuniform rational B-spline interpolation for both the line sources and heart-liver phantom investigated in this study were not found to have the distortion observed from 180° acquisition or a count decrease at the center, resulting in image quality nearly equivalent to 360° acquisition. This method enabled a reduction in testing time without impacting image quality.

OBJECTIVES: Positron emission tomography/computed tomography (PET/CT) examination is commonly used for the evaluation of pulmonary nodules since it provides both anatomical and functional information. However, given the dependence of this evaluation on physician's subjective judgment, the results could be variable. The purpose of this study was to develop an automated scheme for the classification of pulmonary nodules using early and delayed phase PET/CT and conventional CT images. METHODS: We analysed 36 early and delayed phase PET/CT images in patients who underwent both PET/CT scan and lung biopsy, following bronchoscopy. In addition, conventional CT images at maximal inspiration were analysed. The images consisted of 18 malignant and 18 benign nodules. For the classification scheme, 25 types of shape and functional features were first calculated from the images. The random forest algorithm, which is a machine learning technique, was used for classification. RESULTS: The evaluation of the characteristic features and classification accuracy was accomplished using collected images. There was a significant difference between the characteristic features of benign and malignant nodules with regard to standardised uptake value and texture. In terms of classification performance, 94.4% of the malignant nodules were identified correctly assuming that 72.2% of the benign nodules were diagnosed accurately. The accuracy rate of benign nodule detection by means of CT plus two-phase PET images was 44.4% and 11.1% higher than those obtained by CT images alone and CT plus early phase PET images, respectively. CONCLUSION: Based on the findings, the proposed method may be useful to improve the accuracy of malignancy analysis.

Lung cancer is a leading cause of death worldwide. Although computed tomography (CT) examinations are frequently used for lung cancer diagnosis, it can be difficult to distinguish between benign and malignant pulmonary nodules on the basis of CT images alone. Therefore, a bronchoscopic biopsy may be conducted if malignancy is suspected following CT examinations. However, biopsies are highly invasive, and patients with benign nodules may undergo many unnecessary biopsies. To prevent this, an imaging diagnosis with high classification accuracy is essential. In this study, we investigate the automated classification of pulmonary nodules in CT images using a deep convolutional neural network (DCNN). We use generative adversarial networks (GANs) to generate additional images when only small amounts of data are available, which is a common problem in medical research, and evaluate whether the classification accuracy is improved by generating a large amount of new pulmonary nodule images using the GAN. Using the proposed method, CT images of 60 cases with confirmed pathological diagnosis by biopsy are analyzed. The benign nodules assessed in this study are difficult for radiologists to differentiate because they cannot be rejected as being malignant. A volume of interest centered on the pulmonary nodule is extracted from the CT images, and further images are created using axial sections and augmented data. The DCNN is trained using nodule images generated by the GAN and then fine-tuned using the actual nodule images to allow the DCNN to distinguish between benign and malignant nodules. This pretraining and fine-tuning process makes it possible to distinguish 66.7% of benign nodules and 93.9% of malignant nodules. These results indicate that the proposed method improves the classification accuracy by approximately 20% in comparison with training using only the original images.

OBJECTIVE: This study aims to carry out a quantitative analysis with high reproducibility using single-photon emission computed tomography/computed tomography (SPECT/CT); we investigated the optimum parameters for the acquisition and the reconstruction. MATERIALS AND METHODS: SPECT images were acquired with varying time per view using SPECT phantom (JS-10) and the body phantom of National Electrical Manufacturers Association and International Electrotechnical Commission (Body-phantom), respectively. For the image reconstruction condition, we changed the product of subset and iteration (SI product) and the Gaussian filter using a three-dimensional ordered subset expectation maximization. A combination of no scattering correction and no attenuation correction (SC-/AC-) and a combination of scattering correction and attenuation correction by CT images (SC+/AC+) were performed. The dose linearity, the recovery coefficient, the scatter ratio, and the coefficient of variation were evaluated using JS-10. Using Body-phantom, contrast-to-noise ratios of the hot spheres (13, 17 mm) were calculated. Moreover, the change in the maximum standardized uptake value (SUVmax) and the average SUV (SUVmean) were evaluated for each sphere. RESULT: From the evaluation results using the JS-10, dose linearity, recovery coefficient, scatter ratio, and coefficient of variation were all good when time per view was 50-150 s, the Gaussian filter was 8-12 mm, and the SI product was 150. From the evaluation results using Body-phantom, comparing the Gaussian filter with 8 mm and 12 mm, the contrast-to-noise ratio was better for 12 mm and the error rate to the change of the scan-time was up to 3.7%. However, SUVmax and SUVmean using 8 mm were closer to the design value of the phantom. CONCLUSION: It is necessary that Quantitative SPECT be acquired at 50 s or more per view per detection, reconstructed using a three-dimensional ordered subset expectation maximization with SC+/AC+, the SI product is 150 times, and the Gaussian Filter is 8-12 mm. This suggested that the quantitative analysis would be carried out with good reproducibility.

PURPOSE: To develop a method for automated detection of highly integrated sites in SPECT images using bone information obtained from CT images in bone scintigraphy. METHODS: Bone regions on CT images were first extracted, and bones were identified by segmenting multiple regions. Next, regions corresponding to the bone regions on SPECT images were extracted based on the bone regions on CT images. Subsequently, increased uptake regions were extracted from the SPECT image using thresholding and three-dimensional labeling. Last, the ratio of increased uptake regions to all bone regions was calculated and expressed as a quantitative index. To verify the efficacy of this method, a basic assessment was performed using phantom and clinical data. RESULTS: The results of this analytical method using phantoms created by changing the radioactive concentrations indicated that regions of increased uptake were detected regardless of the radioactive concentration. Assessments using clinical data indicated that detection sensitivity for increased uptake regions was 71% and that the correlation between manual measurements and automated measurements was significant (correlation coefficient 0.868). CONCLUSION: These results suggested that automated detection of increased uptake regions on SPECT images using bone information obtained from CT images would be possible.

In a previous study, we developed a hybrid tumor detection method that used both computed tomography (CT) and positron emission tomography (PET) images. However, similar to existing computer-aided detection (CAD) schemes, it was difficult to detect low-contrast lesions that touch to the normal organs such as the chest wall or blood vessels in the lung. In the current study, we proposed a novel lung tumor detection method that uses active contour filters to detect the nodules deemed "difficult" in previous CAD schemes. The proposed scheme detects lung tumors using both CT and PET images. As for the detection in CT images, the massive region was first enhanced using an active contour filter (ACF), which is a type of contrast enhancement filter that has a deformable kernel shape. The kernel shape involves closed curves that are connected by several nodes that move iteratively in order to enclose the massive region. The final output of ACF is the difference between the maximum pixel value on the deformable kernel, and pixel value on the center of the filter kernel. Subsequently, the PET images were binarized to detect the regions of increased uptake. The results were integrated, followed by the false positive reduction using 21 characteristic features and three support vector machines. In the experiment, we evaluated the proposed method using 100 PET/CT images. More than half of nodules missed using previous methods were accurately detected. The results indicate that our method may be useful for the detection of lung tumors using PET/CT images.