Automated Diagnosis of Cardiovascular Disease Through Measurement of Intima Media Thickness Using Deep Neural Networks

Investigating Temporal Features of Carotid Intima-Media Thickness from Ultrasound Imaging with Recurrent Neural Networks

Measuring carotid intima-media thickness (cIMT) of the Common Carotid Artery (CCA) via B-mode ultrasound imaging is a non-invasive yet effective way to monitor and assess cardiovascular risk. Recent studies using Convolutional Neural Networks (CNNs) to automate the process have mainly focused on the detection of regions of interest (ROI) in single frame images collected at fixed time points and have not exploited the temporal information captured in ultrasound imaging. This paper presents a novel framework to investigate the temporal features of cIMT, in which Recurrent Neural Networks (RNN) were deployed for ROI detection using consecutive frames from ultrasound imaging. The cIMT time series can be formed from estimates of cIMT in each frame of an ultrasound scan, from which additional information (such as min, max, mean, and frequency) on cIMT time series can be extracted. Results from evaluation show the best performance for ROI detection improved 4.75% by RNN compared to CNN-based methods. Furthermore, the heart rate estimated from the cIMT time series for seven patients was highly correlated with the patient's clinical records, which suggests the potential application of the cIMT time series and related features for clinical studies in the future.Clinical relevance- The temporal features extracted from cIMT time series provide additional information that can be potentially beneficial for clinical studies.

Measurement of Intima-Media Thickness Depending on Intima Media Complex Segmentation by Deep Neural Networks

Intima Media Thickness (IMT) of the carotid artery is an important marker indicating the sign of cardiovascular disease. Automated measurement of IMT requires segmentation of intima media complex (IMC).Traditional methods which use shape, color and texture for classification have poor generalization capability. This paper proposes two models- the pipeline model and the end-to-end model using Convolutional Neural Networks (CNN) and auto encoder–decoder network respectively. CNN architecture is implemented and tested by varying the number of convolutional layer, size of the kernel as well as the number of kernels. Auto encoder–decoder performs pixel wise classification using two interconnected pathways for identifying the boundary of lumen-intima (LI) and media adventitia (MA). This helps in reconstruction of the segmented portion for measurement of IMT. Both methods are tested using a dataset of 550 subjects. The results clearly indicate that end-to-end model has an edge over the pipeline model exhibiting lesser deviation between the automated measurement and the measurement made by the radiologist. The pipeline model however has better segmentation accuracy when the size of the image used for training is small. The convolutional neural network with auto encoder–decoder proves robust through sparse representation, and faster learning with better generalization. Also, the experimental setup is analyzed by interconnecting Tensor flow simulated result with Raspberry PI and the outcomes are analyzed.

SLAM-OR: Simultaneous Localization, Mapping and Object Recognition Using Video Sensors Data in Open Environments from the Sparse Points Cloud

In this paper, we propose a novel approach that enables simultaneous localization, mapping (SLAM) and objects recognition using visual sensors data in open environments that is capable to work on sparse data point clouds. In the proposed algorithm the ORB-SLAM uses the current and previous monocular visual sensors video frame to determine observer position and to determine a cloud of points that represent objects in the environment, while the deep neural network uses the current frame to detect and recognize objects (OR). In the next step, the sparse point cloud returned from the SLAM algorithm is compared with the area recognized by the OR network. Because each point from the 3D map has its counterpart in the current frame, therefore the filtration of points matching the area recognized by the OR algorithm is performed. The clustering algorithm determines areas in which points are densely distributed in order to detect spatial positions of objects detected by OR. Then by using principal component analysis (PCA)-based heuristic we estimate bounding boxes of detected objects. The image processing pipeline that uses sparse point clouds generated by SLAM in order to determine positions of objects recognized by deep neural network and mentioned PCA heuristic are main novelties of our solution. In contrary to state-of-the-art approaches, our algorithm does not require any additional calculations like generation of dense point clouds for objects positioning, which highly simplifies the task. We have evaluated our research on large benchmark dataset using various state-of-the-art OR architectures (YOLO, MobileNet, RetinaNet) and clustering algorithms (DBSCAN and OPTICS) obtaining promising results. Both our source codes and evaluation data sets are available for download, so our results can be easily reproduced.