Detection of acute rib fractures on CT images with convolutional neural networks: effect of location and type of fracture and reader's experience.

PURPOSE: The evaluation of all ribs on thin-slice CT images is time consuming and it can be difficult to accurately assess the location and type of rib fracture in an emergency. The aim of our study was to develop and validate a convolutional neural network (CNN) algorithm for the detection of acute rib fractures on thoracic CT images and to investigate the effect of the CNN algorithm on radiologists' performance. METHODS: The dataset for development of a CNN consisted of 539 thoracic CT scans with 4906 acute rib fractures. A three-dimensional faster region-based CNN was trained and evaluated by using tenfold cross-validation. For an observer performance study to investigate the effect of CNN outputs on radiologists' performance, 30 thoracic CT scans (28 scans with 90 acute rib fractures and 2 without rib fractures) which were not included in the development dataset were used. Observer performance study involved eight radiologists who evaluated CT images first without and second with CNN outputs. The diagnostic performance was assessed by using figure of merit (FOM) values obtained from the jackknife free-response receiver operating characteristic (JAFROC) analysis. RESULTS: When radiologists used the CNN output for detection of rib fractures, the mean FOM value significantly increased for all readers (0.759 to 0.819, P = 0.0004) and for displaced (0.925 to 0.995, P = 0.0028) and non-displaced fractures (0.678 to 0.732, P = 0.0116). At all rib levels except for the 1st and 12th ribs, the radiologists' true-positive fraction of the detection became significantly increased by using the CNN outputs. CONCLUSION: The CNN specialized for the detection of acute rib fractures on CT images can improve the radiologists' diagnostic performance regardless of the type of fractures and reader's experience. Further studies are needed to clarify the usefulness of the CNN for the detection of acute rib fractures on CT images in actual clinical practice.

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.

[Development of temporal subtraction method for chest radiographs by using pixel matching technique].

A temporal subtraction image, which is obtained by subtraction of a previous image from a current one, can enhance interval changes on a chest radiograph by removal of most normal structures. However, subtraction artifacts, which tend to reduce its effectiveness in the detection of interval changes, were still included in the conventional method. In this study, we have developed a pixel matching technique to reduce artifacts in the temporal subtraction images. With this technique, the pixel value in a nonlinearly warped previous image is replaced by a pixel value within a kernel, which is closest to the pixel value on a current image. For evaluation of the proposed method, one hundred temporal subtraction images with a simulated nodule were used. When the kernel size of 3×3 was employed in the pixel matching technique, the misregistration artifacts decreased by 72%, and the contrast-to-noise ratio of the simulated lung nodules was increased by 5% in comparison with the conventional method. However, the area of the simulated nodule on the subtraction image decreased by 6%. Our results indicated that the pixel matching technique can enhance simulated nodules, with a substantial reduction of misregistration artifacts in comparison with conventional subtraction images. Therefore, we believe that the temporal subtraction method with the pixel matching technique would assist radiologists' diagnoses for detection of lung nodules in digital chest radiography.

[Computer-aided detection of metastatic brain tumors in magnetic resonance images].

The fact that accurate detection of metastatic brain tumors is important for making decisions on the treatment course of patients prompted us to develop a computer-aided diagnostic scheme for detecting metastatic brain tumors. In this paper, we first describe how we extracted the cerebral parenchyma region using a standard deviation filter. Second, initial candidates for tumors were decided by sphericity and cross-correlation value with a simulated ring template. Third, we made true positive and false positive templates obtained from actual clinical images and applied the template matching technique to them. Finally, we detected metastatic tumors using these two characteristics. Our improved method was applied to 13 cases with 97 brain metastases. Sensitivity of detection of metastatic brain tumors was 80.4%, with 5.6 false positives per patient. Our proposed method has potential for detection of metastatic brain tumors in brain magnetic resonance (MR) images.

Development of a voxel-matching technique for substantial reduction of subtraction artifacts in temporal subtraction images obtained from thoracic MDCT.

A temporal subtraction image, which is obtained by subtraction of a previous image from a current one, can be used for enhancing interval changes (such as formation of new lesions and changes in existing abnormalities) on medical images by removing most of the normal structures. However, subtraction artifacts are commonly included in temporal subtraction images obtained from thoracic computed tomography and thus tend to reduce its effectiveness in the detection of pulmonary nodules. In this study, we developed a new method for substantially removing the artifacts on temporal subtraction images of lungs obtained from multiple-detector computed tomography (MDCT) by using a voxel-matching technique. Our new method was examined on 20 clinical cases with MDCT images. With this technique, the voxel value in a warped (or nonwarped) previous image is replaced by a voxel value within a kernel, such as a small cube centered at a given location, which would be closest (identical or nearly equal) to the voxel value in the corresponding location in the current image. With the voxel-matching technique, the correspondence not only between the structures but also between the voxel values in the current and the previous images is determined. To evaluate the usefulness of the voxel-matching technique for removal of subtraction artifacts, the magnitude of artifacts remaining in the temporal subtraction images was examined by use of the full width at half maximum and the sum of a histogram of voxel values, which may indicate the average contrast and the total amount, respectively, of subtraction artifacts. With our new method, subtraction artifacts due to normal structures such as blood vessels were substantially removed on temporal subtraction images. This computerized method can enhance lung nodules on chest MDCT images without disturbing misregistration artifacts.

Semi-automated segmentation of a glioblastoma multiforme on brain MR images for radiotherapy planning.

We propose a computerized method for semi-automated segmentation of the gross tumor volume (GTV) of a glioblastoma multiforme (GBM) on brain MR images for radiotherapy planning (RTP). Three-dimensional (3D) MR images of 28 cases with a GBM were used in this study. First, a sphere volume of interest (VOI) including the GBM was selected by clicking a part of the GBM region in the 3D image. Then, the sphere VOI was transformed to a two-dimensional (2D) image by use of a spiral-scanning technique. We employed active contour models (ACM) to delineate an optimal outline of the GBM in the transformed 2D image. After inverse transform of the optimal outline to the 3D space, a morphological filter was applied to smooth the shape of the 3D segmented region. For evaluation of our computerized method, we compared the computer output with manually segmented regions, which were obtained by a therapeutic radiologist using a manual tracking method. In evaluating our segmentation method, we employed the Jaccard similarity coefficient (JSC) and the true segmentation coefficient (TSC) in volumes between the computer output and the manually segmented region. The mean and standard deviation of JSC and TSC were 74.2+/-9.8% and 84.1+/-7.1%, respectively. Our segmentation method provided a relatively accurate outline for GBM and would be useful for radiotherapy planning.

[A computerized method for detection of acute cerebral infarction on CT images].

This paper proposes a computerized method for automated detection of acute cerebral infarction (ACI) on CT images. This method is based on the difference value of image features in the two regions-of-interests (ROIs) selected at symmetrical positions. In our computerized method, first, we segmented the brain parenchyma by the thresholding technique after correction of inclination of the midsagittal plane with translation and rotation of the image. Then we selected the middle cerebral artery (MCA) region of the brain parenchyma. Moreover, many ROIs with a 32×32 matrix size were selected in the MCA region. In addition, image features in each ROI were determined from the statistical analysis, the co-occurrence matrix and the run length matrix. Finally, ROIs with ACI were classified by using a linear discriminant analysis with difference values of image features in two ROIs at symmetrical positions. Nineteen cases with ACI and normal 14 cases were employed in this study. As a result of our experiments, the sensitivity of detection of ACI was 88.0% with an average number of false positives of 4.6 per case. Our computerized method provided a relatively high performance for detection of ACI. Therefore, we believe this method would be useful for an algorithm of a computer-aided diagnosis to detect ACI on CT images.

Development of a computer simulation technique for low-dose chest radiographs: a phantom study.

The present study aimed to develop a simple computer simulation method of low-dose radiographs based on a radiograph acquired at a clinical-dose level. A chest phantom was used for the development of this method. In this method, a simulated low-dose image was obtained from a clinical-dose image using an input-output characteristic curve of a flat panel detector and noise metrics of the standard deviation (SD) and noise power spectrum. We applied this method for low-dose images of a chest phantom to evaluate the simulation accuracy. The noise SDs were compared between the simulated and real images corresponding to 1/2, 1/4, and 1/8 of clinical doses. The relative error of noise SDs in the chest phantom images was less than 3%. Therefore, we believe that the proposed simulation method has the potential to be useful for determination of the optimal exposure condition in chest radiography to reduce patients' exposure dose.

The effect of deep convolutional neural networks on radiologists' performance in the detection of hip fractures on digital pelvic radiographs.

PURPOSE: The purpose of our study is to develop deep convolutional neural network (DCNN) for detecting hip fractures using CT and MRI as a gold standard, and to evaluate the diagnostic performance of 7 readers with and without DCNN. METHODS: The study population consisted of 327 patients who underwent pelvic CT or MRI and were diagnosed with proximal femoral fractures. All radiographs were manually checked and annotated by radiologists referring to CT and MRI for selecting ROI. At first, a DCNN with the GoogLeNet model was trained by 302 cases. The remaining 25 cases and 25 control subjects were used for the observer performance study and for the testing of DCNN. Seven readers took part in this study. A continuous rating scale was used to record each observer's confidence level. Subsequently, each observer interpreted with the DCNN outputs and rated them again. The area under the curve (AUC) was used to compare the fracture detection. RESULTS: The average AUC of the 7 readers was 0.832. The AUC of DCNN alone was 0.905. The average AUC of the 7 readers with DCNN outputs was 0.876. The AUC of readers with DCNN output were higher than those without(p < 0.05). The AUC of the 2 experienced readers with DCNN output exceeded the AUC of DCNN alone. CONCLUSION: For detecting the hip fractures on radiographs, DCNN developed using CT and MRI as a gold standard by radiologists improved the diagnostic performance including the experienced readers.

Evaluation of the depiction ability of similar subtraction images using digital chest radiographs of different patients.

The temporal subtraction (TS) technique requires the same patient's chest radiographs (CXRs) acquired on different dates, whereas the similar subtraction (SS) technique can be used in patients who have no previous CXR, using similar CXRs from different patients. This study aimed to examine the depiction ability of SS images with simulated nodules in comparison with that of TS images with 2- and 7-year acquisition intervals. One hundred patients were randomly selected from our image database. The most recently acquired images of the patients were used as target images for subtraction. The simulated nodule was superimposed on each target image to examine the usefulness of the SS technique. The most (Top 1) and ten most (Top 10) similar images for each target image were identified in the 24,254-image database using a template-matching technique, and used for the SS technique. SS and TS images were obtained using a previously developed nonlinear image-warping technique. The depiction ability of SS and TS images was evaluated using the contrast-to-noise ratio (CNR). The proportion of Top 1 SS images showing higher CNR than that of the TS images with 2- and 7-year acquisition intervals was 28% (28/100) and 33% (33/100), respectively. Moreover, the proportion of cases that had any of the Top 10 SS images with higher CNRs than those of TS images with 2- and 7-year acquisition intervals was 56% (56/100) and 72% (72/100), respectively. Our study indicates that the SS technique can potentially be used to detect lung nodules on CXRs.

Diagnostic accuracy and reading time to detect intracranial aneurysms on MR angiography using a computer-aided diagnosis system.

OBJECTIVE: The objective of our study was to determine whether the use of a computer-aided diagnosis (CAD) system can shorten the reading time while maintaining the diagnostic performance of MR angiography for the detection of intracranial aneurysms. MATERIALS AND METHODS: Fifty maximum-intensity-projection MR angiograms in 50 patients (16 intracranial aneurysms and 34 negative cases) were used for this observer performance study. Sixteen radiologists--eight neuroradiologists and eight less experienced radiologists--participated in the observer studies and interpreted the MR angiograms without and with CAD output images using an independent test method. The reading times without and with CAD were compared separately for the aneurysm and negative cases. The observers' performances were evaluated using receiver operating characteristic (ROC) analysis. We analyzed separately the data obtained from neuroradiologists and from less experienced radiologists. RESULTS: For all observers, the mean area under the ROC curve (Az) with CAD was improved compared with that without CAD (0.903 vs 0.851, respectively; p = 0.109), and the mean reading time per case was reduced significantly by 18.1 seconds (28.5%) (from 63.4 to 45.3 seconds, p < 0.05). When CAD output images were available, the mean A(z) for the less experienced radiologists was significantly improved (0.911 vs 0.787, p < 0.05), but not for the neuroradiologists. The mean reading time of the less experienced radiologists with CAD was significantly shorter than that of the neuroradiologists without CAD (39.8 vs 54.5 seconds, p < 0.05). CONCLUSION: The use of a CAD system for the detection of intracranial aneurysms on MR angiography can shorten the reading time while improving diagnostic performance for less experienced radiologists.

Improving image quality around subtle lung nodules by reducing artifacts in similar subtraction imaging.

Similar subtraction imaging is useful for the detection of lung nodules; however, some artifacts on similar subtraction images reduce their utility. The authors attempted to improve the image quality of similar subtraction images by reducing artifacts caused by differences in image contrast and sharpness between two images used for similar subtraction imaging. Image contrast was adjusted using the histogram specification technique. The differences in image sharpness were compensated for using a pixel matching technique. The improvement in image quality was evaluated objectively based on the degree of artifacts and the contrast-to-noise ratio (CNR) of the lung nodules. The artifacts in similar subtraction images were reduced in 94% (17/18) of cases, and CNR was improved in 83% (15/18) of cases. The results indicate that the combination of histogram specification and pixel matching techniques is potentially useful in improving image quality in similar subtraction imaging.

Impact of FDG-PET/CT imaging on nodal staging for head-and-neck squamous cell carcinoma.

PURPOSE: To evaluate the impact of 18F-fluorodeoxyglucose positron emission tomography/computed tomography (FDG-PET/CT) imaging on nodal staging for head-and-neck squamous cell carcinoma (SCC). METHODS AND MATERIALS: The study population consisted of 23 patients with head-and-neck SCC who were evaluated with FDG-PET/CT and went on to neck dissection. Two observers consensually determined the lesion size and maximum standardized uptake value (SUVmax) and compared the results with pathologic findings on nodal-level involvement. Two different observers (A and B) independently performed three protocols for clinical nodal staging. Methods 1, 2, and 3 were based on conventional modalities, additional visual information from FDG-PET/CT images, and FDG-PET/CT imaging alone with SUV data, respectively. RESULTS: All primary tumors were visualized with FDG-PET/CT. Pathologically, 19 positive and 93 negative nodal levels were identified. The SUVmax overlapped in negative and positive nodes <15 mm in diameter. According to receiver operating characteristics analysis, the size-based SUVmax cutoff values were 1.9, 2.5, and 3.0 for lymph nodes <10 mm, 10-15 mm, and >15 mm, respectively. These cutoff values yielded 79% sensitivity and 99% specificity for nodal-level staging. For Observer A, the sensitivity and specificity in Methods 1, 2, and 3 were 68% and 94%, 68% and 99%, and 84% and 99%, respectively, and Method 3 yielded significantly higher accuracy than Method 1 (p = 0.0269). For Observer B, Method 3 yielded the highest sensitivity (84%) and specificity (99%); however, the difference among the three protocols was not statistically significant. CONCLUSION: Imaging with FDG-PET/CT with size-based SUVmax cutoff values is an important modality for radiation therapy planning.

Computer-aided diagnosis in chest radiography.

We have developed computer-aided diagnosis (CAD) schemes for the detection of lung nodules, interstitial lung diseases, interval changes, and asymmetric opacities, and also for the differential diagnosis of lung nodules and interstitial lung diseases on chest radiographs. Observer performance studies indicate clearly that radiologists' diagnostic accuracy was improved significantly when radiologists used a computer output in their interpretations of chest radiographs. In addition, the automated recognition methods for the patient and the projection view by use of chest radiographs were useful for integrating the chest CAD schemes into the picture-archiving and communication system (PACS).

Computerized detection of intracranial aneurysms for three-dimensional MR angiography: feature extraction of small protrusions based on a shape-based difference image technique.

We have improved a computerized scheme for the detection of intracranial aneurysms for three-dimensional (3-D) magnetic resonance angiography (MRA) by the use of image features of small protrusions extracted based on a shape-based difference image (SBDI) technique. Initial candidates were identified by use of a multiple gray-level thresholding technique in dot enhanced images, and by finding short branches in skeleton images. Image features related to aneurysms were determined based on candidate regions segmented by use of a region growing technique. For extracting additional features on small protrusions or small aneurysms, we have developed an SBDI technique, which was based on the shape-based difference between an original segmented vessel and a vessel with suppressed local change in thickness. The SBDI technique was useful for obtaining local changes in vessel thickness, i.e., SBD regions, which could be small aneurysms in the case of true positives, but thin or very small regions in the case of false positives. Many false positives were removed by means of rule-based schemes and linear discriminant analysis on various 3-D localized image features, including SBDI features. We tested the computerized scheme on 53 cases with 61 aneurysms and 62 nonaneurysm cases based on a leave-one-out-by-patient test method. As a result, false positives per patient decreased from 5.8 to 3.8, while a high sensitivity of 97% was maintained by use of the SBDI technique, in which SBDI features were effective for removing some false positives. The computer-aided diagnostic (CAD) scheme may be robust and useful in assisting radiologists in the detection of intracranial aneurysms for MRA.