The purpose of this study was to develop a computerized classification method for molecular subtypes of low-grade gliomas (LGGs) in brain MRI (magnetic resonance imaging) images using multi-scale three-dimensional attention branch networks (MS3D-ABNs) with an additive angular margin penalty. Our database consisted of brain T1-weighted, T2-weighted, and FLAIR MRI images for 217 patients (58 IDH-mutant astrocytomas, 49 IDH-wildtype astrocytomas, and 110 oligodendrogliomas). The proposed network was constructed from a feature extractor, an attention branch, and a perception branch with an additive angular margin penalty. The feature extractor first extracted the feature maps of different resolutions from brain T1-weighted, T2-weighted, and FLAIR MRI images, respectively. The attention branch generated attention maps focusing on a tumor region. The feature maps were then multiplied by the attention maps to weight features on the tumor region in the feature maps. The perception branch finally evaluated the molecular subtypes of LGGs by determining the cosine similarity between the feature vector obtained from applying a global average pooling to the feature map and the representative vector of each molecular subtype class. In training the proposed network, an angular margin penalty was added to the angle between the feature vector of input image and the representative vector of the same class as the input image to make those vectors to be closer each other. The classification accuracy for the proposed network was 66.4%, showing an improvement when compared to the MS3D-ABNs without the additive angular margin penalty (60.4%).
Breast magnetic resonance imaging (MRI) has a higher sensitivity of early breast cancer than mammography, but the specificity is lower. In MRI examination at clinical practice, multiple MRI sequences are usually acquired to achieve high diagnostic accuracy. The purpose of this study was to develop a computerized classification scheme for distinguishing between benign and malignant masses by integrally analyzing multiple MRI sequences with convolutional neural networks (CNNs). Our database consisted of four MRI sequences for 43 patients with masses. It included T1-weighted images, T2- weighted images, dynamic contrast material-enhanced magnetic resonance imaging (DCE-MRI) images, and the difference images of the DCE-MRI images for each patient. In training the CNNs, the CNNs were first trained independently for each MRI sequence. The CNN features extracted from four MRI sequences with the trained CNNs were then inputted to a support vector machine (SVM) for distinguishing between benign and malignant masses. A k-fold cross validation method (k=3) was used for training and testing the CNNs and the SVM. With the proposed method, the classification accuracy, the sensitivity, the specificity, the positive predictive value, and the negative predictive value were 88.4% (38/43), 90.0% (27/30), 84.6% (11/13), 78.6% (11/14), and 93.1% (27/29), respectively. The classification performance with the proposed method analyzing multiple MRI sequences was substantially greater than those with CNNs analyzing one MRI sequence. The proposed method achieved high classification performance and would be useful in differential diagnoses of masses as diagnostic aid.
It can be difficult for clinicians to correctly determine biopsy or follow-up for masses on breast ultrasonographic images. The purpose of this study was to develop a computerized determination scheme for histological classification of masses using a combination of CNN (convolutional neural network) features and morphologic features. The database consisted of 585 breast ultrasonographic images. It included 288 malignant masses (218 invasive carcinomas and 70 noninvasive carcinomas) and 297 benign masses (182 fibroadenomas and 115 cysts). In the proposed method, CNN features and morphologic features were first determined from a mass. The CNN features were defined by reducing the dimensionality of the output of the final pooling layer in GoogLeNet using a principal component analysis. The morphologic features were also defined by taking into account image features commonly used for describing masses on breast ultrasonographic images. A support vector machine (SVM) with the CNN features and the morphologic features was employed to classify among histological classifications of masses. Three-fold cross validation method was used for training and testing the GoogLeNet and the SVM. The classification accuracies with the proposed method were 84.4% (184/218) for invasive carcinomas, 72.9% (51/70) for noninvasive carcinomas, 85.7% (156/182) for fibroadenomas, and 87.8% (101/115) for cysts, respectively. The sensitivity and the specificity were 87.2% (251/288) and 93.3% (277/297), whereas the positive predictive value and the negative predictive value were 92.6% (251/271) and 88.2% (277/314). The proposed method yielding high classification accuracies would be useful in the differential diagnosis of masses on ultrasonographic images as diagnosis aid.
Based on the “as low as reasonably practicable” principle, the amount of radiation exposure of CT should be decreased without impairing image quality. However, the excessive reduction of radiation exposure in CT results in the degraded images with noise and artifacts. The purpose of this study was to construct virtual normal dose CT images from ultra-low dose CT images by using dilated residual networks (DRN) that can expand the receptive field. Our database consisted of 1,860 pairs of normal dose and ultra-dose chest CT images obtained from 12 patients. In our proposed method, the DRN which consisted of seven dilated convolutional layers were trained the relationship of signal patterns between ultra-low dose CT patches (small region) and the corresponding normal dose CT patches. The trained DRN was employed to construct the virtual normal dose CT images from the ultra-low dose CT images. The root mean squared error, peak signal to noise ratio, and structural similarity index for the ultra-low dose CT images to the normal dose CT images were 59.3, 32.6dB, and 0.897, whereas those for the constructed images from ultra-low dose CT with a model-based iterative reconstruction (MBIR) were 49.1, 34.1dB, and 0.956. Those indices for virtual normal dose CT images were 44.4, 35.0dB, and 0.964, showing a significant improvement when compared with the ultra-low dose CT images and the constructed images with MBIR. The virtual normal dose CT images achieved higher image quality as compared with the ultra-low dose CT images and the constructed images with MBIR.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.