Fully Convolutional Neural Network Architecture Research Articles

Thunderstorm prediction during pre-tactical air-traffic-flow management using convolutional neural networks

Thunderstorms can be a large source of disruption for European air-traffic management causing a chaotic state of operation within the airspace system. In current practice, air-traffic managers are provided with imprecise forecasts which limit their ability to plan strategically. As a result, weather mitigation is performed using tactical measures with a time horizon of three hours. Increasing the lead time of thunderstorm predictions to the day before operations could help air-traffic managers plan around weather and improve the efficiency of air-traffic-management operations. Emerging techniques based on machine learning have provided promising results, partly attributed to reduced human bias and improved capacity in predicting thunderstorms purely from numerical weather prediction data. In this paper, we expand on our previous work on thunderstorm forecasting, by applying convolutional neural networks (CNNs) to exploit the spatial characteristics embedded in the weather data. The learning task of predicting convection is formulated as a binary-classification problem based on satellite data. The performance of multiple CNN-based architectures, including a fully-convolutional neural network (FCN), a CNN-based encoder–decoder, a U-Net, and a pyramid-scene parsing network (PSPNet) are compared against a multi-layer-perceptron (MLP) network. Our work indicates that CNN-based architectures improve the performance of point-prediction models, with a fully-convolutional neural-network architecture having the best performance. Results show that CNN-based architectures can be used to increase the prediction lead time of thunderstorms. Lastly, a case study illustrating the applications of convection-prediction models in an air-traffic-management setting is presented.

Super-resolution deep neural networks for water classification from free multispectral satellite imagery

Recent years have seen rapid progress in the adoption of fully convolutional neural networks (FCN) to classify optical satellite imagery, made possible by a combination of new FCN architectures, next-generation GPUs, and publicly available satellite imagery from, e.g., the Landsat and Sentinel missions. These satellites offer repeat global coverage at intervals of only a few days at a spatial resolution of ≥10 m, which is sufficient for some but not all applications of interest. A smaller body of literature considers similar tools to classify commercial satellite imagery that offer 1 – 2 orders of magnitude higher spatial resolutions but with limited spatial and temporal coverage. In this work, we develop a super-resolution FCN to achieve the best of both worlds: land cover classification at commercial-level spatial resolutions but with the spatiotemporal coverage of public satellite imagery. To do so, we label 1 – 2 m resolution commercial imagery and use this as training data for super-resolution FCN. As a specific application, we focus on the segmentation of rivers, with the goal of tracking changes in reach-averaged river widths, depths, and discharges over time. We present detailed performance analyses and demonstrate that, surprisingly, we achieve ≳ 90% classification accuracy at meter-scale resolutions from free Sentinel-2 imagery. We extensively validate our model through in situ USGS gage data and ground-truth GPS tracing of river shorelines. By making our super-resolution FCN codes and training weights publicly available, we hope that these tools will be of use to the broader hydrology community and beyond, as the models can be re-trained for other segmentation tasks.

Machine learning enables national assessment of wind plant controls with implications for land use

SummaryAs deployment of wind energy continues to expand, computationally efficient tools for predicting wind plant performance over a wide range of layout designs, technology innovations, and spatial locations are increasingly important for policy and investment decisions. We demonstrate two approaches to training a surrogate model to predict annual energy production (AEP) of parameterized wind plant layouts: one using a Gaussian process (GP) and the other using a fully convolutional neural network (FCNN). We leverage the powerful FCNN architecture by encoding wind plant design parameters and output response surface as an image. The FCNN produces more accurate results than the GP with mean absolute errors equivalent to 1% and 1.9% of plant rated power, respectively, although the GP performs well under limited training data and provides useful uncertainty information. We also evaluate a surrogate model for wake steering, enabling a nationwide assessment of the impact of plant control strategies and plant layout decisions. Across two million locations, we find that wake steering strategies boost AEP with relative gains upwards of 3%. Gains are most pronounced at sites without a dominant wind direction and where layout optimization is less fruitful. Additionally, we perform a nationwide sensitivity analysis showing that wake steering can mitigate wake losses from higher density plant layouts. Our results suggest that regions which have not been previously viable for wind deployment due to moderate wind resources are especially well enhanced by wake steering strategies that could help overcome land constraints and inflexible layout options, potentially identifying new deployment opportunities.

Multi-Res-Attention UNet: A CNN Model for the Segmentation of Focal Cortical Dysplasia Lesions from Magnetic Resonance Images

In this work, we have focused on the segmentation of Focal Cortical Dysplasia (FCD) regions from MRI images. FCD is a congenital malformation of brain development that is considered as the most common causative of intractable epilepsy in adults and children. To our knowledge, the latest work concerning the automatic segmentation of FCD was proposed using a fully convolutional neural network (FCN) model based on UNet. While there is no doubt that the model outperformed conventional image processing techniques by a considerable margin, it suffers from several pitfalls. First, it does not account for the large semantic gap of feature maps passed from the encoder to the decoder layer through the long skip connections. Second, it fails to leverage the salient features that represent complex FCD lesions and suppress most of the irrelevant features in the input sample. We propose Multi-Res-Attention UNet; a novel hybrid skip connection-based FCN architecture that addresses these drawbacks. Moreover, we have trained it from scratch for the detection of FCD from 3T MRI 3D FLAIR images and conducted 5-fold cross-validation to evaluate the model. FCD detection rate (Recall) of 92% was achieved for patient wise analysis.

A New Framework for Performing Cardiac Strain Analysis from Cine MRI Imaging in Mice

Cardiac magnetic resonance (MR) imaging is one of the most rigorous form of imaging to assess cardiac function in vivo. Strain analysis allows comprehensive assessment of diastolic myocardial function, which is not indicated by measuring systolic functional parameters using with a normal cine imaging module. Due to the small heart size in mice, it is not possible to perform proper tagged imaging to assess strain. Here, we developed a novel deep learning approach for automated quantification of strain from cardiac cine MR images. Our framework starts by an accurate localization of the LV blood pool center-point using a fully convolutional neural network (FCN) architecture. Then, a region of interest (ROI) that contains the LV is extracted from all heart sections. The extracted ROIs are used for the segmentation of the LV cavity and myocardium via a novel FCN architecture. For strain analysis, we developed a Laplace-based approach to track the LV wall points by solving the Laplace equation between the LV contours of each two successive image frames over the cardiac cycle. Following tracking, the strain estimation is performed using the Lagrangian-based approach. This new automated system for strain analysis was validated by comparing the outcome of these analysis with the tagged MR images from the same mice. There were no significant differences between the strain data obtained from our algorithm using cine compared to tagged MR imaging. Furthermore, we demonstrated that our new algorithm can determine the strain differences between normal and diseased hearts.

A deep learning-based approach for automatic segmentation and quantification of the left ventricle from cardiac cine MR images

Cardiac MRI has been widely used for noninvasive assessment of cardiac anatomy and function as well as heart diagnosis. The estimation of physiological heart parameters for heart diagnosis essentially require accurate segmentation of the Left ventricle (LV) from cardiac MRI. Therefore, we propose a novel deep learning approach for the automated segmentation and quantification of the LV from cardiac cine MR images. We aim to achieve lower errors for the estimated heart parameters compared to the previous studies by proposing a novel deep learning segmentation method. Our framework starts by an accurate localization of the LV blood pool center-point using a fully convolutional neural network (FCN) architecture called FCN1. Then, a region of interest (ROI) that contains the LV is extracted from all heart sections. The extracted ROIs are used for the segmentation of LV cavity and myocardium via a novel FCN architecture called FCN2. The FCN2 network has several bottleneck layers and uses less memory footprint than conventional architectures such as U-net. Furthermore, a new loss function called radial loss that minimizes the distance between the predicted and true contours of the LV is introduced into our model. Following myocardial segmentation, functional and mass parameters of the LV are estimated. Automated Cardiac Diagnosis Challenge (ACDC-2017) dataset was used to validate our framework, which gave better segmentation, accurate estimation of cardiac parameters, and produced less error compared to other methods applied on the same dataset. Furthermore, we showed that our segmentation approach generalizes well across different datasets by testing its performance on a locally acquired dataset. To sum up, we propose a deep learning approach that can be translated into a clinical tool for heart diagnosis.

Registration-free Face-SSD: Single shot analysis of smiles, facial attributes, and affect in the wild

In this paper, we present a novel single shot face-related task analysis method, called Face-SSD, for detecting faces and for performing various face-related (classification/regression) tasks including smile recognition, face attribute prediction and valence-arousal estimation in the wild. Face-SSD uses a Fully Convolutional Neural Network (FCNN) to detect multiple faces of different sizes and recognise/regress one or more face-related classes. Face-SSD has two parallel branches that share the same low-level filters, one branch dealing with face detection and the other one with face analysis tasks. The outputs of both branches are spatially aligned heatmaps that are produced in parallel—therefore Face-SSD does not require that face detection, facial region extraction, size normalisation, and facial region processing are performed in subsequent steps. Our contributions are threefold: 1) Face-SSD is the first network to perform face analysis without relying on pre-processing such as face detection and registration in advance–Face-SSD is a simple and a single FCNN architecture simultaneously performing face detection and face-related task analysis—those are conventionally treated as separate consecutive tasks; 2) Face-SSD is a generalised architecture that is applicable for various face analysis tasks without modifying the network structure—this is in contrast to designing task-specific architectures; and 3) Face-SSD achieves real-time performance (21 FPS) even when detecting multiple faces and recognising multiple classes in a given image (300 × 300). Experimental results show that Face-SSD achieves state-of-the-art performance in various face analysis tasks by reaching a recognition accuracy of 95.76% for smile detection, 90.29% for attribute prediction, and Root Mean Square (RMS) error of 0.44 and 0.39 for valence and arousal estimation.

Recalibrating Fully Convolutional Networks With Spatial and Channel "Squeeze and Excitation" Blocks.

In a wide range of semantic segmentation tasks, fully convolutional neural networks (F-CNNs) have been successfully leveraged to achieve the state-of-the-art performance. Architectural innovations of F-CNNs have mainly been on improving spatial encoding or network connectivity to aid gradient flow. In this paper, we aim toward an alternate direction of recalibrating the learned feature maps adaptively, boosting meaningful features while suppressing weak ones. The recalibration is achieved by simple computational blocks that can be easily integrated in F-CNNs architectures. We draw our inspiration from the recently proposed "squeeze and excitation" (SE) modules for channel recalibration for image classification. Toward this end, we introduce three variants of SE modules for segmentation: 1) squeezing spatially and exciting channel wise; 2) squeezing channel wise and exciting spatially; and 3) joint spatial and channel SE. We effectively incorporate the proposed SE blocks in three state-of-the-art F-CNNs and demonstrate a consistent improvement of segmentation accuracy on three challenging benchmark datasets. Importantly, SE blocks only lead to a minimal increase in model complexity of about 1.5%, while the Dice score increases by 4%-9% in the case of U-Net. Hence, we believe that SE blocks can be an integral part of future F-CNN architectures.

Fully convolutional multi-scale residual DenseNets for cardiac segmentation and automated cardiac diagnosis using ensemble of classifiers.

Deep fully convolutional neural network (FCN) based architectures have shown great potential in medical image segmentation. However, such architectures usually have millions of parameters and inadequate number of training samples leading to over-fitting and poor generalization. In this paper, we present a novel DenseNet based FCN architecture for cardiac segmentation which is parameter and memory efficient. We propose a novel up-sampling path which incorporates long skip and short-cut connections to overcome the feature map explosion in conventional FCN based architectures. In order to process the input images at multiple scales and view points simultaneously, we propose to incorporate Inception module's parallel structures. We propose a novel dual loss function whose weighting scheme allows to combine advantages of cross-entropy and Dice loss leading to qualitative improvements in segmentation. We demonstrate computational efficacy of incorporating conventional computer vision techniques for region of interest detection in an end-to-end deep learning based segmentation framework. From the segmentation maps we extract clinically relevant cardiac parameters and hand-craft features which reflect the clinical diagnostic analysis and train an ensemble system for cardiac disease classification. We validate our proposed network architecture on three publicly available datasets, namely: (i) Automated Cardiac Diagnosis Challenge (ACDC-2017), (ii) Left Ventricular segmentation challenge (LV-2011), (iii) 2015 Kaggle Data Science Bowl cardiac challenge data. Our approach in ACDC-2017 challenge stood second place for segmentation and first place in automated cardiac disease diagnosis tasks with an accuracy of 100% on a limited testing set (n=50). In the LV-2011 challenge our approach attained 0.74 Jaccard index, which is so far the highest published result in fully automated algorithms. In the Kaggle challenge our approach for LV volume gave a Continuous Ranked Probability Score (CRPS) of 0.0127, which would have placed us tenth in the original challenge. Our approach combined both cardiac segmentation and disease diagnosis into a fully automated framework which is computationally efficient and hence has the potential to be incorporated in computer-aided diagnosis (CAD) tools for clinical application.

Video Object Segmentation without Temporal Information.

Video Object Segmentation, and video processing in general, has been historically dominated by methods that rely on the temporal consistency and redundancy in consecutive video frames. When the temporal smoothness is suddenly broken, such as when an object is occluded, or some frames are missing in a sequence, the result of these methods can deteriorate significantly. This paper explores the orthogonal approach of processing each frame independently, i.e., disregarding the temporal information. In particular, it tackles the task of semi-supervised video object segmentation: the separation of an object from the background in a video, given its mask in the first frame. We present Semantic One-Shot Video Object Segmentation (OSVOS$^\mathrm {S}$S), based on a fully-convolutional neural network architecture that is able to successively transfer generic semantic information, learned on ImageNet, to the task of foreground segmentation, and finally to learning the appearance of a single annotated object of the test sequence (hence one shot). We show that instance-level semantic information, when combined effectively, can dramatically improve the results of our previous method, OSVOS. We perform experiments on two recent single-object video segmentation databases, which show that OSVOS$^\mathrm {S}$S is both the fastest and most accurate method in the state of the art. Experiments on multi-object video segmentation show that OSVOS$^\mathrm {S}$S obtains competitive results.

Quantitative Analysis of Patch-Based Fully Convolutional Neural Networks for Tissue Segmentation on Brain Magnetic Resonance Imaging

Accurate brain tissue segmentation in Magnetic Resonance Imaging (MRI) has attracted the attention of medical doctors and researchers since variations in tissue volume help in diagnosing and monitoring neurological diseases. Several proposals have been designed throughout the years comprising conventional machine learning strategies as well as convolutional neural networks (CNN) approaches. In particular, in this paper, we analyse a sub-group of deep learning methods producing dense predictions. This branch, referred in the literature as Fully CNN (FCNN), is of interest as these architectures can process an input volume in less time than CNNs and local spatial dependencies may be encoded since several voxels are classified at once. Our study focuses on understanding architectural strengths and weaknesses of literature-like approaches. Hence, we implement eight FCNN architectures inspired by robust state-of-the-art methods on brain segmentation related tasks. We evaluate them using the IBSR18, MICCAI2012 and iSeg2017 datasets as they contain infant and adult data and exhibit varied voxel spacing, image quality, number of scans and available imaging modalities. The discussion is driven in three directions: comparison between 2D and 3D approaches, the importance of multiple modalities and overlapping as a sampling strategy for training and testing models. To encourage other researchers to explore the evaluation framework, a public version is accessible to download from our research website.