[Application of Convolutional Neural Network for Evaluating CT Dose Using Image Noise Classification: A Phantom Study].

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.

[Investigation of Clinical Utility of Radiological Technologist's Reading Report as a Second Opinion for Medical Doctor Reading of Digital Mammogram].

PURPOSE: We investigated the clinical utility of a radiological technologist's (RT)'s reports (RRs) as a second opinion by the free-response receiver operating characteristic (FROC) observer study that compared the performance of medical doctors' (MDs') reading of digital mammogram with and without consulting the RR. METHOD: One hundred women (39 malignant, 61 benign or normal) who underwent diagnostic mammography were selected from among 1674 routine clinical images classified by the degree of difficulty and categories for inclusion in the FROC study. The first FROC study performed by three RTs (RT 1-3) was conducted to collect the data for RR utilized in the second FROC study. The second FROC study was performed by five MDs, and the statistical significance of MDs' performances with and without reference to the RR was investigated by figure of merit (FOM). RESULT: The FOM values of three RTs obtained in the first FROC study were 0.529, 0.576, and 0.539, respectively. In the second FROC study, RT 2 had the highest FOM, RT 1 the lowest false positives/case, and RT 3 the highest sensitivity. The average FOM values in the second FROC study for the five MDs with/without reference to the RR were as follows: RT 2's RR was 0.534/0.588 (p=0.003), RT 1's RR was 0.500/0.545 (p=0.099), and RT 3's RR was 0.569/0.592 (p=0.324). CONCLUSION: We concluded that the MDs' performance of reading mammogram was statistically improved by consulting the RR when the RT's reading skill was high.

[Development of Automated Positioning System in General Radiography Examination-Application to Four Directions Cervical Spine Radiography].

To reduce a number of retaking with unnecessary radiation exposure, and to improve a quality of general radiography and efficiency of radiographic procedure, we propose an automated radiographic system that uses reference points on human face for determining exposure angle and beam center. As a preliminary study, we developed an automated positioning method for determining exposure angle and beam center of four directions (4R) cervical spine radiography by using the human body data from 3D reconstructed head-and-neck computed tomography (CT) images. An image for recognizing human-face was reconstructed and used for identifying the reference points. Clinical utility of this proposed method was demonstrated by using "KENZO" to inspect simulated projection X-ray images which were reconstructed from CT volume data. In this study, we reused a huge number of CT images, which were obtained in routine clinical procedure and had archived in medical institutions. This database, therefore, allowed us to develop a new radiological technique without any additional patient dose.

[Image Evaluation with Paired Comparison Method Using Automatic Analysis Software: Comparison of CT Images with Simulated Levels of Exposure Dose].

To simplify a procedure of the observer study with Ura's method of Scheffé's paired comparison and to improve experimental accuracy, we developed a software package to automatically analyze observer study data obtained by using a computer interface developed specially for the ROC observer study. Simulated low-dose CT images were used to demonstrate practical utility of this proposed method with a software package, in terms of a statistical analysis of the change of noise property due to the change of exposure dose. Six radiological technologists were participated in this observer study and compared each of six sample images selected at lower lung and liver slices with dose levels of 100, 80, 60, 40, 20, 10% per case. In the statistical analysis, the average psychological measures were highly correlated with the dose levels (lower lungs: R=0.95, liver: R=0.99). In addition, there were statistically significant differences in all combination of dose levels in liver slices. In conclusion, we demonstrated practical utility of this proposed method in terms of simplification of experimental procedure and the consistency of the analytic results.

[Selection of Radiomic Features for the Classification of Triple-negative Breast Cancer Based on Radiogenomics].

Subtype classification of breast cancer by analyzing the gene expression profile of cancer cells is becoming a standard procedure. Breast cancer subtype classification is more useful than the conventional method because the characteristics of subtype classification is directly connected with the treatment method. However, genetic testing is invasive, and a part of cancer cells may not represent the overall nature of the cancer. In the computer-aided diagnosis (CAD) scheme for differentiation of triple-negative breast cancer (TNBC) by estimating the genetic properties of cancer based on Radiogenomics, principal component analysis (PCA) and least absolute shrinkage and selection operator (Lasso) were used for reducing the dimension of radiomic features, and we compared usefulness of both. We collected 81 magnetic resonance (MR) images, which included 30 TNBC and 51 others, from the public database. From the MR slice images, we selected the slice containing the largest area of the cancer and manually marked the cancer region. We subsequently calculated 294 radiomic features in the cancer region, and reduced the dimension of radiomic features. Finally, linear discriminant analysis, with the dimensionally compressed 10 image features, was used for distinguishing between TNBC and others. Area under the curve (AUC) was 0.60 when we used PCA, whereas AUC was 0.70 when we used Lasso (p=0.0058). Therefore, Lasso is useful for the determination of radiomic features in Radiogenomics.

[Automated Detection of Radiopaque Markers in Chest and Abdomen Radiography].

We aimed to develop a computerized method for the detection of radiopaque markers, such as R and L in chest and abdomen radiography by using the generalized Hough transform and the template matching. To develop the computerized method, we used 200 chest and abdomen images in our institution as training cases. First, two template images for R and L markers were created with the same exposure condition as a chest X-ray. Following various image processing, such as edge detection, thinning and Hough transformed, a look-up table that consisted of distance and direction pairs was built for the generalized Hough transform. All training images were preprocessed with median filter, edge detection, binarization, thinning, back ground removal and labeling. For candidates of markers that were detected as true positive or false positive, their vote and cross-correlation were calculated with the generalized Hough transform. To evaluate this proposed method, a validation test was performed with another database that consisted of 800 chest and abdomen images by use of Mahalanobis distance based on vote and cross-correlation in statistics. The precision of detecting the radiopaque markers for 800 test images was 99.9%. In addition, this method worked out well for some specific images in which markers were overlapped with a human body.

[Quantitation of Cerebral Atrophy due to Normal Aging: Principal Component Analysis with MR Images in Patients' Age Groups].

To evaluate the degree of cerebral atrophy, quantification methods of a difference from a standard normal brain are often used in clinical practice. However, these methods may not evaluate the cerebral atrophy accurately, because they do not take into account any cerebral atrophies due to normal aging. The purpose of this study is to develop a model for taking into account the cerebral atrophy due to normal aging. We obtained 60 normal magnetic resonance (MR) images from the Alzheimer's disease neuroimaging initiative database. These data included 20 images of each age group of 60's, 70's, and 80's, respectively. For anatomical standardization of the images, we used the statistical parametric mapping software and employed a linear grayscale transformation. The principal component (PC) analysis with voxel values of 60 normal MR images was subsequently performed to calculate eigenvectors and PC scores. All cases were projected onto the eigenspace formed by 2nd and 5th PC scores. The experimental result showed separated distributions corresponding to the age groups. In addition, the sites of cerebral atrophy could be recognized by displaying eigenimages. Our proposed method would be useful for the accurate evaluation of cerebral atrophy caused by Alzheimer's disease.

Computer-aided diagnosis for estimating the malignancy grade of hepatocellular carcinoma using contrast-enhanced ultrasound: an ROC observer study.

BACKGROUND & AIMS: We are developing a computer-aided diagnosis (CAD) system for estimating the malignancy grade of hepatocellular carcinoma (HCC) using contrast-enhanced ultrasound (CEUS). In this study, observers estimated the malignancy grade of HCC with and without the cues provided by CAD. MATERIALS AND METHODS: Institutional review board approval was obtained and informed consent was waived. A total of 232 histologically confirmed HCCs were studied: 76 well-differentiated HCC (w-HCC), 133 moderately differentiated HCC (m-HCC), and 23 poorly differentiated HCC (p-HCC). In this observer study, CEUS vascular images acquired using the maximum intensity projection technique were displayed together with static B-mode and Kupffer-phase (defined as 10 min after injection) images. Five hepatologists independently assigned confidence ratings for the malignancy grade of each HCC. Each hepatologist first read each case without CAD and then immediately afterwards with CAD. The observers' rating data were evaluated by multireader multicase receiver operating characteristic (ROC) analysis. RESULTS: The overall sensitivity of our CAD system for discrimination between three histological differentiation grades of HCC was 87.5% (203/232). For discrimination between w-HCC and m/p-HCC, the mean area under the ROC curve (AUC) for the five observers was significantly improved from 0.779 ± 0.074 to 0.872 ± 0.090 with CAD (P = 0.0069). For discrimination between m-HCC and p-HCC, the mean AUC was also significantly improved from 0.713 ± 0.107 to 0.863 ± 0.101 with CAD (P = 0.0321). CONCLUSION: The use of our CAD system can significantly improve the diagnostic performance of hepatologists in discriminating between three histological differentiation grades of HCC using CEUS.

Image Data Mining for Pattern Classification and Visualization of Morphological Changes in Brain MR Images.

Hospital information systems (HISs) and picture archiving and communication systems (PACSs) are archiving large amounts of data (i.e., "big data") that are not being used. Therefore, many research projects in progress are trying to use "big data" for the development of early diagnosis, prediction of disease onset, and personalized therapies. In this study, we propose a new method for image data mining to identify regularities and abnormalities in the large image data sets. We used 70 archived magnetic resonance (MR) images that were acquired using three-dimensional magnetization-prepared rapid acquisition with gradient echo (3D MP-RAGE). These images were obtained from the Alzheimer's disease neuroimaging initiative (ADNI) database. For anatomical standardization of the data, we used the statistical parametric mapping (SPM) software. Using a similarity matrix based on cross-correlation coefficients (CCs) calculated from an anatomical region and a hierarchical clustering technique, we classified all the abnormal cases into five groups. The Z score map identified the difference between a standard normal brain and each of those from the Alzheimer's groups. In addition, the scatter plot obtained from two similarity matrixes visualized the regularities and abnormalities in the image data sets. Image features identified using our method could be useful for understanding of image findings associated with Alzheimer's disease.

Eigenspace template matching for detection of lacunar infarcts on MR images.

Detection of lacunar infarcts is important because their presence indicates an increased risk of severe cerebral infarction. However, accurate identification is often hindered by the difficulty in distinguishing between lacunar infarcts and enlarged Virchow-Robin spaces. Therefore, we developed a computer-aided detection (CAD) scheme for the detection of lacunar infarcts. Although our previous CAD method indicated a sensitivity of 96.8% with 0.71 false positives (FPs) per slice, further reduction of FPs remained an issue for the clinical application. Thus, the purpose of this study is to improve our CAD scheme by using template matching in the eigenspace. Conventional template matching is useful for the reduction of FPs, but it has the following two pitfalls: (1) It needs to maintain a large number of templates to improve the detection performance, and (2) calculation of the cross-correlation coefficient with these templates is time consuming. To solve these problems, we used template matching in the lower dimension space made by a principal component analysis. Our database comprised 1,143 T1- and T2-weighted images obtained from 132 patients. The proposed method was evaluated by using twofold cross-validation. By using this method, 34.1% of FPs was eliminated compared with our previous method. The final performance indicated that the sensitivity of the detection of lacunar infarcts was 96.8% with 0.47 FPs per slice. Therefore, the modified CAD scheme could improve FP rate without a significant reduction in the true positive rate.

Incident reports related to tasks performed by radiological technologists: an analysis of ten years of incident reports.

In the past 10 years at our university hospital, 202 incident reports related to tasks performed by radiological technologists were posted. In order to investigate the causes and trends of these incidents, we classified the incident reports into four groups based on the event content, level of harm caused to the patient, years of experience of the concerned radiological technologist, and relevant departmental section. In the event content group, 'a malfunctioning device' was the most common event (26.2%), whereas the other events were 'wrong examination procedure or therapy' (15.3%), 'patient fall' (10.9%), 'procedure-patient mismatch' (8.4%), 'accidental removal of patients' tubes or other intravenous devices' (7.9%), and 'bringing metallic material into the magnetic resonance imaging (MRI) room' (7.4%). In the level of harm caused to the patient group, level one events occurred frequently. Radiological technologists with 6-16 years of experience reported incidents most frequently. With regard to the relevant departmental section where the incidents occurred, departments with the highest number of reports were ranked as follows in descending order: general X-ray examination section, MRI section, radiation therapy section, nuclear medicine (NM) section, computed tomography (CT) section, angiography section, and fluoroscopy section. The following events in each corresponding section require careful monitoring: patient fall in the general X-ray examination section and NM section, bringing metallic material into the MRI room, malfunctioning devices in the radiation therapy section, accidental removal of the patient's tubes in the CT section, incorrect handling of the automatic contrast medium injector in the angiography section, and damage of device or article in the fluoroscopy section.

Temporal subtraction method for lung nodule detection on successive thoracic CT soft-copy images.

PURPOSE: To assess the effects of a new computed tomographic (CT) temporal subtraction (TS) method on radiologist performance in lung nodule detection on thin-section CT images. MATERIALS AND METHODS: The institutional review board approved this study, and the informed consent requirement was waived. Fifty pairs (current and previous CT images) of standard-dose 2-mm thin-section CT images and corresponding CT TS images were used for an observer performance study. Two thoracic radiologists identified 30 nodules ranging in size from 5 to 19 mm, and these nodules served as the reference standard of actionable nodules (noncalcified nodules larger than 4 mm). Eight radiologists (four attending radiologists, four radiology residents) participated in this observer study. Ratings and locations of lesions determined by observers were used to assess the significance of differences between radiologists' performances without and with the CT TS images in jacknife free-response receiver operating characteristics analysis. RESULTS: Average figure of merit values increased significantly for all radiologists (from 0.838 without CT TS images to 0.894 with CT TS images [P = .033]). Average sensitivity for detection of actionable nodules was improved from 73.4% to 83.4%, with a false-positive rate of 0.15 per case, by using CT TS images. The reading time with CT TS images was not significantly different from that without. CONCLUSION: The novel CT TS method would increase observer performance for lung nodule detection without considerably extending the reading time.

Improved detection of cholesterol gallstones using quasi-material decomposition images generated from single-energy computed tomography images via deep learning.

In this study, we developed a method for generating quasi-material decomposition (quasi-MD) images from single-energy computed tomography (SECT) images using a deep convolutional neural network (DCNN). Our aim was to improve the detection of cholesterol gallstones and to determine the clinical utility of quasi-MD images. Four thousand pairs of virtual monochromatic images (70 keV) and MD images (fat/water) of the same section, obtained via dual-energy computed tomography (DECT), were used to train the DCNN. The trained DCNN can automatically generate quasi-MD images from the SECT images. Additional SECT images were obtained from 70 patients (40 with and 30 without cholesterol gallstones) to generate quasi-MD images for testing. The presence of gallstones in this dataset was confirmed by ultrasonography. We conducted a receiver operating characteristic (ROC) observer study with three radiologists to validate the clinical utility of the quasi-MD images for detecting cholesterol gallstones. The mean area under the ROC curve for the detection of cholesterol gallstones improved from 0.867 to 0.921 (p = 0.001) when quasi-MD images were added to SECT images. The clinical utility of quasi-MD imaging for detecting cholesterol gallstones was showed. This study demonstrated that the lesion detection capability of images obtained from SECT can be improved using a DCNN trained with DECT images obtained using high-end computed tomography systems.

Task-based assessment of resolution properties of CT images with a new index using deep convolutional neural network.

In this study, we propose a method for obtaining a new index to evaluate the resolution properties of computed tomography (CT) images in a task-based manner. This method applies a deep convolutional neural network (DCNN) machine learning system trained on CT images with known modulation transfer function (MTF) values to output an index representing the resolution properties of the input CT image [i.e., the resolution property index (RPI)]. Sample CT images were obtained for training and testing of the DCNN by scanning the American Radiological Society phantom. Subsequently, the images were reconstructed using a filtered back projection algorithm with different reconstruction kernels. The circular edge method was used to measure the MTF values, which were used as teacher information for the DCNN. The resolution properties of the sample CT images used to train the DCNN were created by intentionally varying the field of view (FOV). Four FOV settings were considered. The results of adapting this method to the filtered back projection (FBP) and hybrid iterative reconstruction (h-IR) images indicated highly correlated values with the MTF10% in both cases. Furthermore, we demonstrated that the RPIs could be estimated in the same manner under the same imaging conditions and reconstruction kernels, even for other CT systems, where the DCNN was trained on CT systems produced by the same manufacturer. In conclusion, the RPI, which is a new index that represents the resolution property using the proposed method, can be used to evaluate the resolution of a CT system in a task-based manner.