Limited Training Set Research Articles

Satellite-based estimation of monthly mean hourly 1-km urban air temperature using a diurnal temperature cycle model

Cities worldwide face escalating climate change risks, underscoring the need for spatially and temporally resolved urban air temperature (Ta) data. While satellite-derived land surface temperature (LST) data have been widely used to estimate Ta, high-resolution hourly Ta estimation in urban areas remains underexplored. Traditional methods typically rely on LST data from geostationary satellites and continuous 24-h Ta observations from weather stations. To address these limitations, we introduce a method that combines a diurnal temperature cycle (DTC) model with a random forest model to estimate monthly mean hourly urban Ta at 1-km resolution. This approach leverages a limited number of diurnal Ta observations from weather stations, MODIS LST data, and ancillary information. The core idea of the proposed method is to transform the estimation of monthly mean hourly 1-km Ta into estimating 1-km DTC model parameters, primarily daily maximum and minimum Ta values. This method capitalizes on MODIS LST's ability to estimate daily Ta extremes and requires only four diurnal Ta observations within a daily cycle to estimate monthly mean hourly 1-km Ta. Station-based five-fold cross-validation yields overall RMSE values consistently below 1.0 °C across nine cities with diverse geographic and climatic contexts. The accuracy achieved with only four diurnal Ta observations rivals that obtained using continuous 24-h Ta observations. Even with a limited training set of ten stations, the overall RMSE remains below 1.0 °C for most cities. The proposed method proves effective for both single-city and multi-city modeling and can estimate daily hourly 1-km Ta under clear-sky conditions. In conclusion, this study offers a feasible, efficient, and versatile method for accurately estimating monthly mean hourly 1-km Ta, which can be readily applied to other cities and holds potential for various applications.

Multi-scale deep learning ensemble for segmentation of endometriotic lesions

Ultrasound is a readily available, non-invasive and low-cost screening for the identification of endometriosis lesions, but its diagnostic specificity strongly depends on the experience of the operator. For this reason, computer-aided diagnosis tools based on Artificial Intelligence techniques can provide significant help to the clinical staff, both in terms of workload reduction and in increasing the overall accuracy of this type of examination and its outcome. However, although these techniques are spreading rapidly in a variety of domains, their application to endometriosis is still very limited. To fill this gap, we propose and evaluate a novel multi-scale ensemble approach for the automatic segmentation of endometriosis lesions from transvaginal ultrasounds. The peculiarity of the method lies in its high discrimination capability, obtained by combining, in a fusion fashion, multiple Convolutional Neural Networks trained on data at different granularity. The experimental validation carried out shows that: (i) the proposed method allows to significantly improve the performance of the individual neural networks, even in the presence of a limited training set; (ii) with a Dice coefficient of 82%, it represents a valid solution to increase the diagnostic efficacy of the ultrasound examination against such a pathology.

Brave the Wind and the Waves: Discovering Robust and Generalizable Graph Lottery Tickets.

The training and inference of Graph Neural Networks (GNNs) are costly when scaling up to large-scale graphs. Graph Lottery Ticket (GLT) has presented the first attempt to accelerate GNN inference on large-scale graphs by jointly pruning the graph structure and the model weights. Though promising, GLT encounters robustness and generalization issues when deployed in real-world scenarios, which are also long-standing and critical problems in deep learning ideology. In real-world scenarios, the distribution of unseen test data is typically diverse. We attribute the failures on out-of-distribution (OOD) data to the incapability of discerning causal patterns, which remain stable amidst distribution shifts. In traditional spase graph learning, the model performance deteriorates dramatically as the graph/network sparsity exceeds a certain high level. Worse still, the pruned GNNs are hard to generalize to unseen graph data due to limited training set at hand. To tackle these issues, we propose the Resilient Graph Lottery Ticket (RGLT) to find more robust and generalizable GLT in GNNs. Concretely, we reactivate a fraction of weights/edges by instantaneous gradient information at each pruning point. After sufficient pruning, we conduct environmental interventions to extrapolate potential test distribution. Finally, we perform last several rounds of model averages to further improve generalization. We provide multiple examples and theoretical analyses that underpin the universality and reliability of our proposal. Further, RGLT has been experimentally verified across various independent identically distributed (IID) and out-of-distribution (OOD) graph benchmarks.

AE-NeRF: Audio Enhanced Neural Radiance Field for Few Shot Talking Head Synthesis

Audio-driven talking head synthesis is a promising topic with wide applications in digital human, film making and virtual reality. Recent NeRF-based approaches have shown superiority in quality and fidelity compared to previous studies. However, when it comes to few-shot talking head generation, a practical scenario where only few seconds of talking video is available for one identity, two limitations emerge: 1) they either have no base model, which serves as a facial prior for fast convergence, or ignore the importance of audio when building the prior; 2) most of them overlook the degree of correlation between different face regions and audio, e.g., mouth is audio related, while ear is audio independent. In this paper, we present Audio Enhanced Neural Radiance Field (AE-NeRF) to tackle the above issues, which can generate realistic portraits of a new speaker with few-shot dataset. Specifically, we introduce an Audio Aware Aggregation module into the feature fusion stage of the reference scheme, where the weight is determined by the similarity of audio between reference and target image. Then, an Audio-Aligned Face Generation strategy is proposed to model the audio related and audio independent regions respectively, with a dual-NeRF framework. Extensive experiments have shown AE-NeRF surpasses the state-of-the-art on image fidelity, audio-lip synchronization, and generalization ability, even in limited training set or training iterations.

MP2RAGE vs. MPRAGE surface-based morphometry in focal epilepsy.

In drug-resistant focal epilepsy, detecting epileptogenic lesions using MRI poses a critical diagnostic challenge. Here, we assessed the utility of MP2RAGE-a T1-weighted sequence with self-bias correcting properties commonly utilized in ultra-high field MRI-for the detection of epileptogenic lesions using a surface-based morphometry pipeline based on FreeSurfer, and compared it to the common approach using T1w MPRAGE, both at 3T. We included data from 32 patients with focal epilepsy (5 MRI-positive, 27 MRI-negative with lobar seizure onset hypotheses) and 94 healthy controls from two epilepsy centres. Surface-based morphological measures and intensities were extracted and evaluated in univariate GLM analyses as well as multivariate unsupervised 'novelty detection' machine learning procedures. The resulting prediction maps were analyzed over a range of possible thresholds using alternative free-response receiver operating characteristic (AFROC) methodology with respect to the concordance with predefined lesion labels or hypotheses on epileptogenic zone location. We found that MP2RAGE performs at least comparable to MPRAGE and that especially analysis of MP2RAGE image intensities may provide additional diagnostic information. Secondly, we demonstrate that unsupervised novelty-detection machine learning approaches may be useful for the detection of epileptogenic lesions (maximum AFROC AUC 0.58) when there is only a limited lesional training set available. Third, we propose a statistical method of assessing lesion localization performance in MRI-negative patients with lobar hypotheses of the epileptogenic zone based on simulation of a random guessing process as null hypothesis. Based on our findings, it appears worthwhile to study similar surface-based morphometry approaches in ultra-high field MRI (≥ 7 T).

Nosocomial transmission of Aspergillus flavus in a neonatal intensive care unit: Long-term persistence in environment and interest of MALDI-ToF mass-spectrometry coupled with convolutional neural network for rapid clone recognition.

Aspergillosis of the newborn remains a rare but severe disease. We report four cases of primary cutaneous Aspergillus flavus infections in premature newborns linked to incubators contamination by putative clonal strains. Our objective was to evaluate the ability of matrix-assisted laser desorption/ionisation time of flight (MALDI-TOF) coupled to convolutional neural network (CNN) for clone recognition in a context where only a very small number of strains are available for machine learning. Clinical and environmental A. flavus isolates (n=64) were studied, 15 were epidemiologically related to the four cases. All strains were typed using microsatellite length polymorphism. We found a common genotype for 9/15 related strains. The isolates of this common genotype were selected to obtain a training dataset (6 clonal isolates/25 non-clonal) and a test dataset (3 clonal isolates/31 non-clonal), and spectra were analysed with a simple CNN model. On the test dataset using CNN model, all 31 non-clonal isolates were correctly classified, 2/3 clonal isolates were unambiguously correctly classified, whereas the third strain was undetermined (i.e., the CNN model was unable to discriminate between GT8 and non-GT8). Clonal strains of A. flavus have persisted in the neonatal intensive care unit for several years. Indeed, two strains of A. flavus isolated from incubators in September 2007 are identical to the strain responsible for the second case that occurred 3 years later. MALDI-TOF is a promising tool for detecting clonal isolates of A. flavus using CNN even with a limited training set for limited cost and handling time.

KGSR: A kernel guided network for real-world blind super-resolution

In recent years, deep learning-based methods have emerged as dominant players in the field of super-resolution (SR), owing to their exceptional reconstruction performance. The primary driver of their effectiveness lies in their utilization of extensive sets of paired low-resolution and high-resolution images for training deep learning models. This training enables the models to effectively replicate the intricate mapping relationship between low-resolution and high-resolution images. Nevertheless, at present, acquiring a sufficient quantity of such image pairs that satisfy the requirements remains a formidable obstacle. Therefore, in order to break the restriction of limited training sets, self-supervised learning has been introduced to train a model for each low-quality image, without requiring pairwise ground-truths. However, they generally presuppose the generation of low-resolution (LR) images from their high-resolution (HR) counterparts using a pre-defined kernel, such as Bicubic downscaling. Such an assumption is seldom valid for real-world LR images, where degradation processes in practical applications are diverse, intricate, and often undisclosed. Therefore, when the presumed downscaling kernel does not match the actual one, the outcomes of state-of-the-art approaches degrade substantially. In this paper, we introduce KGSR, a kernel-guided network for addressing real-world blind SR, effectively avoiding requiring large training image pairs and transforming the blind image super-resolution problem into a supervised learning and non-blind scenario. Specifically, KGSR trains two networks, namely Upscaling and Downscaling, utilizing only patches extracted from the input test image. On one hand, owing to the cross-scale recurrence property of the SR kernel within a single image, the Downscaling network acquires knowledge of the image-specific degradation process through a generative adversarial network. Consequently, the Downscaling network is capable of generating a downsampled version of the LR test image even when the acquisition process is unknown or less than ideal. Additionally, we employ a dedicated discriminator to compel the Downscaling network to prioritize the characterization of kernel orientations. Conversely, a precise blur kernel has the potential to yield superior performance. Guided by the accurate image-specific SR kernel acquired from the Downscaling network and the downsampled LR input, the Upscaling network is capable of producing a high-quality HR image from the LR input. Within the Upscaling network, we additionally introduce an effective module for harnessing the acquired image-specific SR kernel. KGSR operates as a fully unsupervised approach, yet it can concurrently produce both the image-specific SR kernel and high-quality HR images. Comprehensive experiments conducted on standard benchmarks validate the efficacy of the proposed approach compared to state-of-the-art methodologies. Moreover, the suggested method can deliver visually appealing SR outcomes while exhibiting shorter processing times when applied to real-world LR images.

Left Ventricle Contouring in Cardiac Images in the Internet of Medical Things via Deep Reinforcement Learning

Assessment of the left ventricle segmentation in cardiac magnetic resonance imaging (MRI) is of crucial importance for cardiac disease diagnosis. However, conventional manual segmentation is a tedious task that requires excessive human effort, which makes automated segmentation highly desirable in practice to facilitate the process of clinical diagnosis. The internet of medical things (IoMT) and artificial intelligence (AI) for efficient medical data collection and analysis have been deemed effective approaches to remote and automatic diagnosis. In this paper, we propose a novel reinforcement-learning-based framework for left ventricle contouring, which mimics how a cardiologist outlines the left ventricle in a cardiac image. Since such a contour drawing process is simply moving a paintbrush along a specific trajectory, it is thus analogized to a path finding problem. Following the algorithm of proximal policy optimization (PPO), we train a policy network, which makes a stochastic decision on the agents movement according to its local observation such that the generated trajectory matches the true contour of the left ventricle as much as possible. Moreover, we design a deep learning model with a customized loss function to generate the agents landing spot (or coordinate of its initial position on a cardiac image). We further propose an alternative approach for generating the landing spot based on interventricular septum detection, which is more efficient since no extra effort in data preprocessing and model training are involved. The experiment results show that the coordinates of the generated landing spots with both of the two approaches are sufficiently close to the true contour and the proposed reinforcement-learning-based approach outperforms the existing U-net model and its improved version, even with limited training set.

A contrastive learning framework for safety information extraction in construction

Typically named entity recognition (NER) and relation extraction (RE) from safety documentation (e.g., accident reports) adopt a pipeline processing approach whereby tasks are split into two sub-tasks. As a result, error propagation occurs between components, and useful information from one task may go unexploited by the other. Additionally, training sets to perform NER and RE from safety documentation are often limited and context-specific. Thus, our research addresses the following question: How can we accurately identify entities and extract relations from safety documentation using limited training sets? This paper utilizes ‘contrastive learning’ to tackle our research question. It proposes a contrastive learning-based cascade binary tagging framework (CasRel) to automatically and synchronously extract entities and relations from safety documents. A five-fold cross-validation process is used to validate the effectiveness and feasibility of our approach. Results from the validation process achieve an average precision of 77.8%, recall of 58.7%, and F1-score of 66.9%, outperforming CasRel with an increase of about 10% in precision, 5% in recall, and 7% in F1-score. Thus, our approach can accurately recognize entities and extract relations from safety documentation. The contributions of our study are twofold: (1) an improved unified model is developed to recognize and extract the entity and relation from safety documents to reduce error propagation and improve its accuracy; and (2) the concept of ‘contrastive learning’ is introduced in the design of the joint entity and relation extraction model with limited training sets.

Accelerating the characterization of dynamic DNA origami devices with deep neural networks

Mechanical characterization of dynamic DNA nanodevices is essential to facilitate their use in applications like molecular diagnostics, force sensing, and nanorobotics that rely on device reconfiguration and interactions with other materials. A common approach to evaluate the mechanical properties of dynamic DNA nanodevices is by quantifying conformational distributions, where the magnitude of fluctuations correlates to the stiffness. This is generally carried out through manual measurement from experimental images, which is a tedious process and a critical bottleneck in the characterization pipeline. While many tools support the analysis of static molecular structures, there is a need for tools to facilitate the rapid characterization of dynamic DNA devices that undergo large conformational fluctuations. Here, we develop a data processing pipeline based on Deep Neural Networks (DNNs) to address this problem. The YOLOv5 and Resnet50 network architecture were used for the two key subtasks: particle detection and pose (i.e. conformation) estimation. We demonstrate effective network performance (F1 score 0.85 in particle detection) and good agreement with experimental distributions with limited user input and small training sets (~ 5 to 10 images). We also demonstrate this pipeline can be applied to multiple nanodevices, providing a robust approach for the rapid characterization of dynamic DNA devices.

Feature-level consistency regularized Semi-supervised scheme with data augmentation for intelligent fault diagnosis under small samples

Intelligent fault diagnosis based on machine learning has yielded a wealth of research results. However, fault diagnosis under small samples is still a challenging problem due to the difficulty of collecting fault data of machines in engineering scenarios. To address this problem, this paper proposes a feature-level consistency regularized semi-supervised scheme with data augmentation for fault diagnosis of machines. In the proposed method, generative adversarial networks generate unlabeled data to augment the limited training set. The fault identification network in the proposed method is trained by the augmented data and a small amount of real data collected from machines in a semi-supervised learning manner. We construct a distance evaluation metric based on the Earth Mover's Distance (EMD) and further design a novel feature consistency regularization module, which helps the fault identification network learn robust fault features by minimizing the EMD between generated data features and real data features. To verify the effectiveness of the proposed model, two mechanical fault simulation experiments were carried out. The experimental results show that the proposed method achieves high fault diagnosis accuracy using only a small amount of training data samples.

AttOmics: attention-based architecture for diagnosis and prognosis from omics data.

The increasing availability of high-throughput omics data allows for considering a new medicine centered on individual patients. Precision medicine relies on exploiting these high-throughput data with machine-learning models, especially the ones based on deep-learning approaches, to improve diagnosis. Due to the high-dimensional small-sample nature of omics data, current deep-learning models end up with many parameters and have to be fitted with a limited training set. Furthermore, interactions between molecular entities inside an omics profile are not patient specific but are the same for all patients. In this article, we propose AttOmics, a new deep-learning architecture based on the self-attention mechanism. First, we decompose each omics profile into a set of groups, where each group contains related features. Then, by applying the self-attention mechanism to the set of groups, we can capture the different interactions specific to a patient. The results of different experiments carried out in this article show that our model can accurately predict the phenotype of a patient with fewer parameters than deep neural networks. Visualizing the attention maps can provide new insights into the essential groups for a particular phenotype. The code and data are available at https://forge.ibisc.univ-evry.fr/abeaude/AttOmics. TCGA data can be downloaded from the Genomic Data Commons Data Portal.

Applicability of deep learning optical flow estimation for PIV methods

Recently, there have been some attempts to introduce the method of deep learning-based optical flow(DL-based OF) into PIV measurements. These attempts have made it possible to estimate the flow field with higher accuracy and higher spatial resolution. The current mainstream deep-learning optical flow estimation method often requires a training set whose data distribution is closer to the real sample when completing the target task. However, there is a certain difference between the rigid body motion dataset commonly used in the field of computer vision and the fluid motion dataset in PIV measurements. Whether networks trained with rigid-body data can be directly applied to fluids has not been fully discussed. However, in existing attempts to introduce the DL-based OF method into the field of PIV measurement, the influence of the characteristics of the PIV particle image on the calculation results of the deep learning optical flow estimation (DL-based OF) has not been fully discussed. Therefore, the boundary conditions of the learning method for PIV calculation cannot be fully demonstrated. In this study, we selected the Flownet2 network, which is a classical baseline model on rigid body datasets, as a representative deep learning optical flow estimation method for experiments. Based on the characteristics of various particle images, this study generates a dataset with a wide range of parameters, such as particle density, particle diameter, sheet thickness, and maximum displacement, which are key to the generation of PIV simulation images and discusses the sensitivity of DL-based OF methods under extreme parameter conditions. This study compares (1) the model parameters trained using rigid body datasets such as Sintel, (2) the model parameters trained using synthetic PIV images, (3) the PIV-WIDIM method, and (4) the Optical Flow-Farneback method. The results reveal that the DL-based OF methods in (1) and (2) have a higher spatial resolution and estimation accuracy than the traditional methods in (3)] and (4). Training with the synthetic fluid dataset results in a smoother flow field in method (2)] compared with training with the rigid body dataset in method (1). In the case of limited training set parameters, Flownet2 can still be suitable for extreme conditions, such as small particle concentration, small particle diameter, insufficient particle brightness, or small particle displacement.

Deep learning-based method for segmenting epithelial layer of tubules in histopathological images of testicular tissue.

There is growing concern that male reproduction is affected by environmental chemicals. One way to determine the adverse effect of environmental pollutants is to use wild animals as monitors and evaluate testicular toxicity using histopathology. We propose an automated method to process histology images of testicular tissue. Testicular tissue consists of seminiferous tubules. Segmenting the epithelial layer of the seminiferous tubule is a prerequisite for developing automated methods to detect abnormalities in tissue. We suggest an encoder-decoder fully connected convolutional neural network model to segment the epithelial layer of the seminiferous tubules in histological images. The ResNet-34 is used in the feature encoder module, and the squeeze and excitation attention block is integrated into the encoding module improving the segmentation and localization of epithelium. We applied the proposed method for the two-class problem, where the epithelial layer of the tubule is the target class. The -score and Intersection over Union of the proposed method are 0.85 and 0.92. Although the proposed method is trained on a limited training set, it performs well on an independent dataset and outperforms other state-of-the-art methods. The pretrained ResNet-34 in the encoder and attention block suggested in the decoder result in better segmentation and generalization. The proposed method can be applied to testicular tissue images from any mammalian species and can be used as the first part of a fully automated testicular tissue processing pipeline. The dataset and codes are publicly available on GitHub.

Deep learning-based method for segmenting epithelial layer of tubules in histopathological images of testicular tissue.

There is growing concern that male reproduction is affected by environmental chemicals. One way to determine the adverse effect of environmental pollutants is to use wild animals as monitors and evaluate testicular toxicity using histopathology. We propose an automated method to process histology images of testicular tissue. Testicular tissue consists of seminiferous tubules. Segmenting the epithelial layer of the seminiferous tubule is a prerequisite for developing automated methods to detect abnormalities in tissue. We suggest an encoder-decoder fully connected convolutional neural network model to segment the epithelial layer of the seminiferous tubules in histological images. The ResNet-34 is used in the feature encoder module, and the squeeze and excitation attention block is integrated into the encoding module improving the segmentation and localization of epithelium. We applied the proposed method for the two-class problem, where the epithelial layer of the tubule is the target class. The -score and Intersection over Union of the proposed method are 0.85 and 0.92. Although the proposed method is trained on a limited training set, it performs well on an independent dataset and outperforms other state-of-the-art methods. The pretrained ResNet-34 in the encoder and attention block suggested in the decoder result in better segmentation and generalization. The proposed method can be applied to testicular tissue images from any mammalian species and can be used as the first part of a fully automated testicular tissue processing pipeline. The dataset and codes are publicly available on GitHub.

Translating AI to Clinical Practice: Overcoming Data Shift with Explainability.

To translate artificial intelligence (AI) algorithms into clinical practice requires generalizability of models to real-world data. One of the main obstacles to generalizability is data shift, a data distribution mismatch between model training and real environments. Explainable AI techniques offer tools to detect and mitigate the data shift problem and develop reliable AI for clinical practice. Most medical AI is trained with datasets gathered from limited environments, such as restricted disease populations and center-dependent acquisition conditions. The data shift that commonly exists in the limited training set often causes a significant performance decrease in the deployment environment. To develop a medical application, it is important to detect potential data shift and its impact on clinical translation. During AI training stages, from premodel analysis to in-model and post hoc explanations, explainability can play a key role in detecting model susceptibility to data shift, which is otherwise hidden because the test data have the same biased distribution as the training data. Performance-based model assessments cannot effectively distinguish the model overfitting to training data bias without enriched test sets from external environments. In the absence of such external data, explainability techniques can aid in translating AI to clinical practice as a tool to detect and mitigate potential failures due to data shift. ©RSNA, 2023 Quiz questions for this article are available in the supplemental material.

Data Preprocessing and Neural Network Architecture Selection Algorithms in Cases of Limited Training Sets—On an Example of Diagnosing Alzheimer’s Disease

This paper aimed to increase accuracy of an Alzheimer’s disease diagnosing function that was obtained in a previous study devoted to application of decision roots to the diagnosis of Alzheimer’s disease. The obtained decision root is a discrete switching function of several variables applicated to aggregation of a few indicators to one integrated assessment presents as a superposition of few functions of two variables. Magnetic susceptibility values of the basal veins and veins of the thalamus were used as indicators. Two categories of patients were used as function values. To increase accuracy, the idea of using artificial neural networks was suggested, but a feature of medical data is its limitation. Therefore, neural networks based on limited training datasets may be inefficient. The solution to this problem is proposed to preprocess initial datasets to determine the parameters of the neural networks based on decisions’ roots, because it is known that any can be represented in the incompletely connected neural network form with a cascade structure. There are no publicly available specialized software products allowing the user to set the complex structure of a neural network, which is why the number of synaptic coefficients of an incompletely connected neural network has been determined. This made it possible to predefine fully connected neural networks, comparable in terms of the number of unknown parameters. Acceptable accuracy was obtained in cases of one-layer and two-layer fully connected neural networks trained on limited training sets on an example of diagnosing Alzheimer’s disease. Thus, the scientific hypothesis on preprocessing initial datasets and neural network architecture selection using special methods and algorithms was confirmed.

Guided Wave Damage Location of Pressure Vessel Based on Optimized Explainable Convolutional Neural Network for Multivariate Time Series Classification Neural Network

Abstract Guided wave is a key nondestructive technique for structural health monitoring due to its high sensitivity to structural changes and long propagation distance. However, to achieve high accuracy for damage location, large quantities of samples and thousands of iterations are typically needed for detection algorithms. To address this, in this paper, an eXplainable Convolutional neural network for Multivariate time series classification (XCM) is adopted, which is composed of one-dimensional (1D) and two-dimensional (2D) convolution layers to achieve high accuracy damage location on pressure vessels with limited training sets. By further optimizing the network parameters and network structure, the training time is greatly reduced and the accuracy is further improved. The optimized XCM improves the damage location precision from 95.5% to 98% with small samples (training set/validation set/testing set = 23/2/25) and low training epochs (under 100 epochs), suggesting that the XCM has great advantages in pressure vessel's damage location classification its potential for guided wave-based damage detection techniques in structural health monitoring.

Placenta segmentation in ultrasound imaging: Addressing sources of uncertainty and limited field-of-view.

Automatic segmentation of the placenta in fetal ultrasound (US) is challenging due to the (i) high diversity of placenta appearance, (ii) the restricted quality in US resulting in highly variable reference annotations, and (iii) the limited field-of-view of US prohibiting whole placenta assessment at late gestation. In this work, we address these three challenges with a multi-task learning approach that combines the classification of placental location (e.g., anterior, posterior) and semantic placenta segmentation in a single convolutional neural network. Through the classification task the model can learn from larger and more diverse datasets while improving the accuracy of the segmentation task in particular in limited training set conditions. With this approach we investigate the variability in annotations from multiple raters and show that our automatic segmentations (Dice of 0.86 for anterior and 0.83 for posterior placentas) achieve human-level performance as compared to intra- and inter-observer variability. Lastly, our approach can deliver whole placenta segmentation using a multi-view US acquisition pipeline consisting of three stages: multi-probe image acquisition, image fusion and image segmentation. This results in high quality segmentation of larger structures such as the placenta in US with reduced image artifacts which are beyond the field-of-view of single probes.

New multiple sclerosis lesion segmentation and detection using pre-activation U-Net.

Automated segmentation of new multiple sclerosis (MS) lesions in 3D MRI data is an essential prerequisite for monitoring and quantifying MS progression. Manual delineation of such lesions is time-consuming and expensive, especially because raters need to deal with 3D images and several modalities. In this paper, we propose Pre-U-Net, a 3D encoder-decoder architecture with pre-activation residual blocks, for the segmentation and detection of new MS lesions. Due to the limited training set and the class imbalance problem, we apply intensive data augmentation and use deep supervision to train our models effectively. Following the same U-shaped architecture but different blocks, Pre-U-Net outperforms U-Net and Res-U-Net on the MSSEG-2 dataset, achieving a Dice score of 40.3% on new lesion segmentation and an F1 score of 48.1% on new lesion detection. The codes and trained models are publicly available at https://github.com/pashtari/xunet.