Interval Cancer Detection Using a Neural Network and Breast Density in Women with Negative Screening Mammograms.

Background Inclusion of mammographic breast density (BD) in breast cancer risk models improves accuracy, but accuracy remains modest. Interval cancer (IC) risk prediction may be improved by combining assessments of BD and an artificial intelligence (AI) cancer detection system. Purpose To evaluate the performance of a neural network (NN)-based model that combines the assessments of BD and an AI system in the prediction of risk of developing IC among women with negative screening mammography results. Materials and Methods This retrospective nested case-control study performed with screening examinations included women who developed IC and women with normal follow-up findings (from January 2011 to January 2015). An AI cancer detection system analyzed all studies yielding a score of 1-10, representing increasing likelihood of malignancy. BD was automatically computed using publicly available software. An NN model was trained by combining the AI score and BD using 10-fold cross-validation. Bootstrap analysis was used to calculate the area under the receiver operating characteristic curve (AUC), sensitivity at 90% specificity, and 95% CIs of the AI, BD, and NN models. Results A total of 2222 women with IC and 4661 women in the control group were included (mean age, 61 years; age range, 49-76 years). AUC of the NN model was 0.79 (95% CI: 0.77,0.81), which was higher than AUC of the AI cancer detection system or BD alone (AUC, 0.73 [95% CI: 0.71, 0.76] and 0.69 [95% CI: 0.67, 0.71], respectively; P < .001 for both). At 90% specificity, the NN model had a sensitivity of 50.9% (339 of 666 women; 95% CI: 45.2, 56.3) for prediction of IC, which was higher than that of the AI system (37.5%; 250 of 666 women; 95% CI: 33.0, 43.7; P < .001) or BD percentage alone (22.4%; 149 of 666 women; 95% CI: 17.9, 28.5; P < .001). Conclusion The combined assessment of an artificial intelligence detection system and breast density measurements enabled identification of a larger proportion of women who would develop interval cancer compared with either method alone. Published under a CC BY 4.0 license.

Fully automated detection of breast cancer in screening MRI using convolutional neural networks.

Current computer-aided detection (CADe) systems for contrast-enhanced breast MRI rely on both spatial information obtained from the early-phase and temporal information obtained from the late-phase of the contrast enhancement. However, late-phase information might not be available in a screening setting, such as in abbreviated MRI protocols, where acquisition is limited to early-phase scans. We used deep learning to develop a CADe system that exploits the spatial information obtained from the early-phase scans. This system uses three-dimensional (3-D) morphological information in the candidate locations and the symmetry information arising from the enhancement differences of the two breasts. We compared the proposed system to a previously developed system, which uses the full dynamic breast MRI protocol. For training and testing, we used 385 MRI scans, containing 161 malignant lesions. Performance was measured by averaging the sensitivity values between 1/8-eight false positives. In our experiments, the proposed system obtained a significantly ([Formula: see text]) higher average sensitivity ([Formula: see text]) compared with that of the previous CADe system ([Formula: see text]). In conclusion, we developed a CADe system that is able to exploit the spatial information obtained from the early-phase scans and can be used in screening programs where abbreviated MRI protocols are used.

Textural analysis of early-phase spatiotemporal changes in contrast enhancement of breast lesions imaged with an ultrafast DCE-MRI protocol.

PURPOSE: New ultrafast view-sharing sequences have enabled breast dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) to be performed at high spatial and temporal resolution. The aim of this study is to evaluate the diagnostic potential of textural features that quantify the spatiotemporal changes of the contrast-agent uptake in computer-aided diagnosis of malignant and benign breast lesions imaged with high spatial and temporal resolution DCE-MRI. METHOD: The proposed approach is based on the textural analysis quantifying the spatial variation of six dynamic features of the early-phase contrast-agent uptake of a lesion's largest cross-sectional area. The textural analysis is performed by means of the second-order gray-level co-occurrence matrix, gray-level run-length matrix and gray-level difference matrix. This yields 35 textural features to quantify the spatial variation of each of the six dynamic features, providing a feature set of 210 features in total. The proposed feature set is evaluated based on receiver operating characteristic (ROC) curve analysis in a cross-validation scheme for random forests (RF) and two support vector machine classifiers, with linear and radial basis function (RBF) kernel. Evaluation is done on a dataset with 154 breast lesions (83 malignant and 71 benign) and compared to a previous approach based on 3D morphological features and the average and standard deviation of the same dynamic features over the entire lesion volume as well as their average for the smaller region of the strongest uptake rate. RESULT: The area under the ROC curve (AUC) obtained by the proposed approach with the RF classifier was 0.8997, which was significantly higher (P = 0.0198) than the performance achieved by the previous approach (AUC = 0.8704) on the same dataset. Similarly, the proposed approach obtained a significantly higher result for both SVM classifiers with RBF (P = 0.0096) and linear kernel (P = 0.0417) obtaining AUC of 0.8876 and 0.8548, respectively, compared to AUC values of previous approach of 0.8562 and 0.8311, respectively. CONCLUSION: The proposed approach based on 2D textural features quantifying spatiotemporal changes of the contrast-agent uptake significantly outperforms the previous approach based on 3D morphology and dynamic analysis in differentiating the malignant and benign breast lesions, showing its potential to aid clinical decision making.

Using deep learning to segment breast and fibroglandular tissue in MRI volumes.

PURPOSE: Automated segmentation of breast and fibroglandular tissue (FGT) is required for various computer-aided applications of breast MRI. Traditional image analysis and computer vision techniques, such atlas, template matching, or, edge and surface detection, have been applied to solve this task. However, applicability of these methods is usually limited by the characteristics of the images used in the study datasets, while breast MRI varies with respect to the different MRI protocols used, in addition to the variability in breast shapes. All this variability, in addition to various MRI artifacts, makes it a challenging task to develop a robust breast and FGT segmentation method using traditional approaches. Therefore, in this study, we investigated the use of a deep-learning approach known as "U-net." MATERIALS AND METHODS: We used a dataset of 66 breast MRI's randomly selected from our scientific archive, which includes five different MRI acquisition protocols and breasts from four breast density categories in a balanced distribution. To prepare reference segmentations, we manually segmented breast and FGT for all images using an in-house developed workstation. We experimented with the application of U-net in two different ways for breast and FGT segmentation. In the first method, following the same pipeline used in traditional approaches, we trained two consecutive (2C) U-nets: first for segmenting the breast in the whole MRI volume and the second for segmenting FGT inside the segmented breast. In the second method, we used a single 3-class (3C) U-net, which performs both tasks simultaneously by segmenting the volume into three regions: nonbreast, fat inside the breast, and FGT inside the breast. For comparison, we applied two existing and published methods to our dataset: an atlas-based method and a sheetness-based method. We used Dice Similarity Coefficient (DSC) to measure the performances of the automated methods, with respect to the manual segmentations. Additionally, we computed Pearson's correlation between the breast density values computed based on manual and automated segmentations. RESULTS: The average DSC values for breast segmentation were 0.933, 0.944, 0.863, and 0.848 obtained from 3C U-net, 2C U-nets, atlas-based method, and sheetness-based method, respectively. The average DSC values for FGT segmentation obtained from 3C U-net, 2C U-nets, and atlas-based methods were 0.850, 0.811, and 0.671, respectively. The correlation between breast density values based on 3C U-net and manual segmentations was 0.974. This value was significantly higher than 0.957 as obtained from 2C U-nets (P < 0.0001, Steiger's Z-test with Bonferoni correction) and 0.938 as obtained from atlas-based method (P = 0.0016). CONCLUSIONS: In conclusion, we applied a deep-learning method, U-net, for segmenting breast and FGT in MRI in a dataset that includes a variety of MRI protocols and breast densities. Our results showed that U-net-based methods significantly outperformed the existing algorithms and resulted in significantly more accurate breast density computation.

A computer-aided diagnosis system for breast DCE-MRI at high spatiotemporal resolution.

PURPOSE: With novel MRI sequences, high spatiotemporal resolution has become available in dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) of the breast. Since benign structures in the breast can show enhancement similar to malignancies in DCE-MRI, characterization of detected lesions is an important problem. The purpose of this study is to develop a computer-aided diagnosis (CADx) system for characterization of breast lesions imaged with high spatiotemporal resolution DCE-MRI. METHODS: The developed CADx system is composed of four main parts: semiautomated lesion segmentation, automated computation of morphological and dynamic features, aorta detection, and classification between benign and malignant categories. Lesion segmentation is performed by using a "multiseed smart opening" algorithm. Five morphological features were computed based on the segmentation of the lesion. For each voxel, contrast enhancement curve was fitted to an exponential model and dynamic features were computed based on this fitted curve. Average and standard deviations of the dynamic features were computed over the entire segmented area, in addition to the average value in an automatically selected smaller "most suspicious region." To compute the dynamic features for an enhancement curve, information of aortic enhancement is also needed. To keep the system fully automated, the authors developed a component which automatically detects the aorta and computes the aortic enhancement time. The authors used random forests algorithm to classify benign lesions from malignant. The authors evaluated this system in a dataset of breast MRI scans of 325 patients with 223 malignant and 172 benign lesions and compared its performance to an existing approach. The authors also evaluated the classification performances for ductal carcinoma in situ (DCIS), invasive ductal carcinoma (IDC), and invasive lobular carcinoma (ILC) lesions separately. The classification performances were measured by receiver operating characteristic (ROC) analysis in a leave-one-out cross validation scheme. RESULTS: The area under the ROC curve (AUC) obtained by the proposed CADx system was 0.8543, which was significantly higher (p = 0.007) than the performance obtained by the previous CADx system (0.8172) on the same dataset. The AUC values for DCIS, IDC, and ILC lesions were 0.7924, 0.8688, and 0.8650, respectively. CONCLUSIONS: The authors developed a CADx system for high spatiotemporal resolution DCE-MRI of the breast. This system outperforms a previously proposed system in classifying benign and malignant lesions, while it requires less user interactions.

Similarity analysis of functional connectivity with functional near-infrared spectroscopy.

One of the remaining challenges in functional connectivity (FC) studies is investigation of the temporal variability of FC networks. Recent studies focusing on the dynamic FC mostly use functional magnetic resonance imaging as an imaging tool to investigate the temporal variability of FC. We attempted to quantify this variability via analyzing the functional near-infrared spectroscopy (fNIRS) signals, which were recorded from the prefrontal cortex (PFC) of 12 healthy subjects during a Stroop test. Mutual information was used as a metric to determine functional connectivity between PFC regions. Two-dimensional correlation based similarity measure was used as a method to analyze within-subject and intersubject consistency of FC maps and how they change in time. We found that within-subject consistency (0.61±0.09 ) is higher than intersubject consistency (0.28±0.13 ). Within-subject consistency was not found to be task-specific. Results also revealed that there is a gradual change in FC patterns during a Stroop session for congruent and neutral conditions, where there is no such trend in the presence of an interference effect. In conclusion, we have demonstrated the between-subject, within-subject, and temporal variability of FC and the feasibility of using fNIRS for studying dynamic FC.