Convolutional Encoder-decoder Architecture Research Articles

Multi scale encoder-decoder network with Time Frequency Attention and S-TCN for single channel speech enhancement

The goal of speech enhancement is to restore clean speech in noisy environments. Acoustic scenarios with low signal-to-noise ratios (SNR) make it quite challenging to extract the target speech from its noise. In the current study, to enhance noisy speech, we propose a feature recalibration based multi-scale convolutional encoder-decoder architecture with squeeze temporal convolutional networks (S-TCN) bottleneck. Each multi-scale convolutional layer in encoder and decoder is followed by time-frequency attention module (TFA). The recalibration based multi-scale 2D convolution layers are used to extract local and contextual information. Additionally, the recalibration network is equipped with a gating mechanism to control the flow of information among the layers, enabling weighting of the scaled features for noise suppression and speech retention. The fully connected layer (FC) in the bottleneck part of encoder-decoder contains a few neurons, which capture the global information from the multi-scale 2D convolution layer and reduce parameters. A S-TCN, inspired by the popular temporal convolutional neural network (TCNN), is inserted between the encoder and the decoder to model long-term dependencies in speech. The TFA is a highly efficient network component, that operates through two simultaneous attentions, one focused on time frames, and the other on frequency channels. These attentions work together to explicitly exploit positional information to create a two-dimensional attention map to effectively capture the significant time-frequency distribution of speech. Utilizing the common voice dataset, our proposed model consistently enhances results compared to the current benchmarks, as demonstrated by two extensively utilized objective measures PESQ and STOI. The proposed model shows significant improvements, with average PESQ and STOI scores increasing by 45.7% and 23.8% respectively for seen background noises, and by 43.5% and 21.4% for unseen background noises, when compared to the quality of noisy speech. Tests validate that the proposed approach outperforms numerous cutting-edge algorithms.

A new real-time image semantic segmentation framework based on a lightweight deep convolutional encoder-decoder architecture for robotic environment sensing

Background: Image semantic segmentation can be understood as the allocation of a predefined category label to each pixel in the image to achieve the region segmentation of the image. Different categories in the image are identified with different colors. While achieving pixel classification, the position information of pixel points of different categories in the image is retained. Purpose: Due to the influence of background and complex environment, the traditional semantic segmentation methods have low accuracy. To alleviate the above problems, this paper proposes a new real-time image semantic segmentation framework based on a lightweight deep convolutional encoder-decoder architecture for robotic environment sensing. Methodology: This new framework is divided into three stages: encoding stage, decoding stage and dimension reduction stage. In the coding stage, a cross-layer feature map fusion (CLFMF) method is proposed to improve the effect of feature extraction. In the decoding stage, a new lightweight decoder (LD) structure is designed to reduce the number of convolutional layers to speed up model training and prediction. In the dimension reduction stage, the convolution dimension reduction method (CDR) is presented to connect the encoder and decoder layer by layer to enhance the decoder effect. Results: Compared with other state-of-the-art image semantic segmentation methods, we conduct comparison experiments on datasets Cityscapes, SUN RGB-D, CamVid, KITTI. The Category iIoU combined with the proposed method is more than 70%, and the Category IoU is as high as 89.7%. Conclusion: The results reflect that the new method can achieve the better semantic segmentation effect.

A surrogate modeling method for distributed land surface hydrological models based on deep learning

Surrogate modelling methods have been used to estimate spatially distributed parameters of land surface hydrological models due to their computational efficiency. However, traditional surrogate techniques have trouble in modelling high-dimensional input–output maps. In addition, their applications to a distributed model over large spatial and temporal domains are very unlikely to produce consistent spatial pattern and dynamic relationship of model input and output, leading to unbalanced model performance across different spatiotemporal domains. To address these issues, we employ a deep autoregressive neural network to construct an accurate and reliable surrogate system of distributed dynamic model outputs. The network uses a deep convolutional encoder-decoder architecture to take full advantage of the deep convolutional layers in high-dimensional image processing and spatial feature extraction. Moreover, the autoregressive strategy makes the network good at handling dynamic relationship between the time-varying model input and output. We apply this deep learning (DL) network to building a surrogate model of the Variable Infiltration Capacity (VIC) model simulated soil moisture over the Huaihe river basin of China. The results show that the adopted deep autoregressive neural network can provide an accurate surrogate model of the high-dimensional dynamic relationship between 15 input fields and 1 output field, each covering 408 grids, over multiple years. This surrogate model significantly outperforms the popular long short-term memory (LSTM) model, achieving an average mean squared error (MSE) of 0.26 and an average R2 of 0.76 using the test datasets. The surrogate's MSE and R2 performance represents an average improvement of 43% and 33%, respectively, compared to the LSTM model. In addition, it shows better spatial performance than the LSTM model, demonstrating its superior performance in approximating spatial patterns of VIC simulated soil moisture. The accuracy of this surrogate method and its ability to handle high-dimensional data is promising for advancing parameter uncertainty quantification of distributed land surface hydrological models, which usually have high dimensionality due to the large number of variables involved, including distributed parameters, input forcing variables, and output variables.

Multitasking neural network to jointly map discrete fracture structures and matrix transmissivity by inverting hydraulic data acquired in 2D fractured aquifers. XNET-fracture

In this paper, we propose a new approach to simultaneously map the fracture/conduit network and the equivalent transmissivity of the rock matrix from hydraulic head measurements acquired during pumping/injection tests in fractured aquifers. The algorithm relies on the use of a multitask neural network to directly approximate the joint inversion operator that links hydraulic head data to aquifer hydraulic properties. In which, the hydraulic head responses are used as input data while the output consists of the fracture structure and matrix transmissivity field. The multitasking is formed by fusing two single-task neural networks, both based on convolutional encoder-decoder architectures, one of which processes the fracture map inversion and the other the transmissivity map. Training neural network then relies on a large database, which consists of thousands of synthetic aquifers characterized by the presence of the fracture/conduit network and a heterogeneous transmissivity field attributed to the rock matrix, both randomly generated. In these aquifer models, the groundwater flow equation is solved in a discrete-continuum concept to numerically simulate the hydraulic head responses associated with sequential pumping/injection tests.Multitasking succeeds in reconstructing the fracture architecture and matrix transmissivity field with higher accuracy in comparison to conventional single-task networks. The difference is due to the transfer mechanism involved in the multi-task network, where the information shared between the tasks effectively enhances the accuracy of both reconstructions. However, as with other inversion techniques, the result accuracy depends on the quality and quantity of the hydraulic head data used in the inversion, as well as the prior information involved in featuring the training models.

Encoder-Decoder Architecture for 3D Seismic Inversion.

Inverting seismic data to build 3D geological structures is a challenging task due to the overwhelming amount of acquired seismic data, and the very-high computational load due to iterative numerical solutions of the wave equation, as required by industry-standard tools such as Full Waveform Inversion (FWI). For example, in an area with surface dimensions of 4.5 km × 4.5 km, hundreds of seismic shot-gather cubes are required for 3D model reconstruction, leading to Terabytes of recorded data. This paper presents a deep learning solution for the reconstruction of realistic 3D models in the presence of field noise recorded in seismic surveys. We implement and analyze a convolutional encoder-decoder architecture that efficiently processes the entire collection of hundreds of seismic shot-gather cubes. The proposed solution demonstrates that realistic 3D models can be reconstructed with a structural similarity index measure (SSIM) of 0.9143 (out of 1.0) in the presence of field noise at 10 dB signal-to-noise ratio.

Unsupervised segmentation of biomedical hyperspectral image data: tackling high dimensionality with convolutional autoencoders.

Information about the structure and composition of biopsy specimens can assist in disease monitoring and diagnosis. In principle, this can be acquired from Raman and infrared (IR) hyperspectral images (HSIs) that encode information about how a sample's constituent molecules are arranged in space. Each tissue section/component is defined by a unique combination of spatial and spectral features, but given the high dimensionality of HSI datasets, extracting and utilising them to segment images is non-trivial. Here, we show how networks based on deep convolutional autoencoders (CAEs) can perform this task in an end-to-end fashion by first detecting and compressing relevant features from patches of the HSI into low-dimensional latent vectors, and then performing a clustering step that groups patches containing similar spatio-spectral features together. We showcase the advantages of using this end-to-end spatio-spectral segmentation approach compared to i) the same spatio-spectral technique not trained in an end-to-end manner, and ii) a method that only utilises spectral features (spectral k-means) using simulated HSIs of porcine tissue as test examples. Secondly, we describe the potential advantages/limitations of using three different CAE architectures: a generic 2D CAE, a generic 3D CAE, and a 2D convolutional encoder-decoder architecture inspired by the recently proposed UwU-net that is specialised for extracting features from HSI data. We assess their performance on IR HSIs of real colon samples. We find that all architectures are capable of producing segmentations that show good correspondence with HE stained adjacent tissue slices used as approximate ground truths, indicating the robustness of the CAE-driven spatio-spectral clustering approach for segmenting biomedical HSI data. Additionally, we stress the need for more accurate ground truth information to enable a precise comparison of the advantages offered by each architecture.

A deep image-to-image network organ segmentation algorithm for radiation treatment planning: principles and evaluation

BackgroundWe describe and evaluate a deep network algorithm which automatically contours organs at risk in the thorax and pelvis on computed tomography (CT) images for radiation treatment planning.MethodsThe algorithm identifies the region of interest (ROI) automatically by detecting anatomical landmarks around the specific organs using a deep reinforcement learning technique. The segmentation is restricted to this ROI and performed by a deep image-to-image network (DI2IN) based on a convolutional encoder-decoder architecture combined with multi-level feature concatenation. The algorithm is commercially available in the medical products “syngo.via RT Image Suite VB50” and “AI-Rad Companion Organs RT VA20” (Siemens Healthineers). For evaluation, thoracic CT images of 237 patients and pelvic CT images of 102 patients were manually contoured following the Radiation Therapy Oncology Group (RTOG) guidelines and compared to the DI2IN results using metrics for volume, overlap and distance, e.g., Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD95). The contours were also compared visually slice by slice.ResultsWe observed high correlations between automatic and manual contours. The best results were obtained for the lungs (DSC 0.97, HD95 2.7 mm/2.9 mm for left/right lung), followed by heart (DSC 0.92, HD95 4.4 mm), bladder (DSC 0.88, HD95 6.7 mm) and rectum (DSC 0.79, HD95 10.8 mm). Visual inspection showed excellent agreements with some exceptions for heart and rectum.ConclusionsThe DI2IN algorithm automatically generated contours for organs at risk close to those by a human expert, making the contouring step in radiation treatment planning simpler and faster. Few cases still required manual corrections, mainly for heart and rectum.

Multi-Task SE-Network for Image Splicing Localization

Image splicing can be easily used for illegal activities such as falsifying propaganda for political purposes and reporting false news, which may result in negative impacts on society. Hence, it is highly required to detect spliced images and localize the spliced regions. In this work, we propose a multi-task squeeze and excitation network (SE-Network) for splicing localization. The proposed network consists of two streams, namely label mask stream and edge-guided stream, both of which adopt convolutional encoder-decoder architecture. The information from the edge-guided stream is transmitted to the label mask stream for enhancing the discrimination of features between the spliced and host regions. This work has three main contributions. First, image edges, along with label masks and mask edges, are exploited to supply more comprehensive supervision for the localization of spliced regions. Second, the low-level feature maps extracted from shallow layers are fused with the high-level feature maps from deep layers to provide more reliable feature for splicing localization. Finally, several squeeze and excitation attention modules are incorporated into the network to recalibrate the fused features to enhance the feature expression. Extensive experiments show that the proposed multi-task SE-Network outperforms existing splicing localization methods evidently on two synthetic splicing datasets and four benchmark splicing datasets.

A temporal-spectral generative adversarial network based end-to-end packet loss concealment for wideband speech transmission.

Packet loss concealment (PLC) aims to mitigate speech impairments caused by packet losses so as to improve speech perceptual quality. This paper proposes an end-to-end PLC algorithm with a time-frequency hybrid generative adversarial network, which incorporates a dilated residual convolution and the integration of a time-domain discriminator and frequency-domain discriminator into a convolutional encoder-decoder architecture. The dilated residual convolution is employed to aggregate the short-term and long-term context information of lost speech frames through two network receptive fields with different dilation rates, and the integrated time-frequency discriminators are proposed to learn multi-resolution time-frequency features from correctly received speech frames with both time-domain waveform and frequency-domain complex spectrums. Both causal and noncausal strategies are proposed for the packet-loss problem, which can effectively reduce the transitional distortion caused by lost speech frames with a significantly reduced number of training parameters and computational complexity. The experimental results show that the proposed method can achieve better performance in terms of three objective measurements, including the signal-to-noise ratio, perceptual evaluation of speech quality, and short-time objective intelligibility. The results of the subjective listening test further confirm a better performance in the speech perceptual quality.

Use of Deep Learning for Weather Radar Nowcasting

AbstractNowcasting based on weather radar uses the current and past observations to make estimations of future radar echoes. There are many types of operationally deployed nowcasting systems, but none of them are currently based on deep learning, despite it being an active area of research in the last few years. This paper explores deep learning models as alternatives to current methods by proposing different architectures and comparing them against some operational nowcasting systems. The methods proposed here, harnessing residual convolutional encoder-decoder architectures, reach a level of performance expected of current systems and in certain scenarios can even outperform them. Finally, some of the potential drawbacks of using deep learning are analyzed. No decay in the performance on a different geographical area from where the models were trained was found. No edge or checkerboard artifact, common in convolutional operations, was found that affects the nowcasting metrics.

A deep convolutional encoder-decoder architecture for autonomous fault detection of PV plants using multi-copters

This study presents an autonomous fault detection method for a wide range of common failures and defects which are visually visible on PV modules. In this paper, we focus especially on detection of bird’s drops as a very typical defect on the PV modules. As a crucial prerequisite, a data-set of aerial imageries of the PV strings affected by bird’s drops were collected through several experimental flight by multi-copters in order to train an accurate fully convolutional deep network. These images are divided into three groups, namely, training, testing, and validation parts. For the purpose of bird’s drops segmentation, an improved encoder-decoder architecture is employed. In this regard, a modified VGG16 model is used as a backbone for the encoder part. The encoder of the network has a very flexible architecture that can be modified and trained for any other visual failure detection. Later on, extracted feature maps of images are imported into a decoder network to map the low resolution features to full resolution ones for pixel-wise segmentation. In addition, an image object positioning algorithm is presented to find the exact position of detected failures in local coordinate system. In a post-processing step, the detected damages are prioritized based on various parameters such as severity of shading and extent of impact on the PV module’s output current. For further validation, different affected PV modules were characterized according to the output patterns of the classification step in order to accurately evaluate the effect of birds’ drops and consequent shading on the parameters of PV modules based on their severity and location. Finally, the training and testing results demonstrate that the proposed FCN network is able to predict precisely covered pixels by bird’s drops on PV modules at pixel level with average accuracies of 98% and 93% for training and testing, respectively.

Sequential Fault Detection and Classification in Chemical Process Using a Deep Convolutional Encoder-Decoder Architecture Based on System Dynamics

In this paper, we propose a high-level spatiotemporal feature extraction based on deep convolutional bidirectional encoder-decoder representation network with gated recurrent unit (GRU) cell for dynamic fault diagnosis. Although traditional recurrent neural networks (RNNs) have a major issue with extracting the spatial dependencies, temporal convolutional operation is designed to extract the local features which provide insight into different types of faults. The system dynamics are further extracted with the context by the bidirectional encoder-decoder network, which distils the representations from future time steps. Moreover, GRU cell in both encoder and decoder deploys a more compact structure than a usual RNNs network due to the reduction of gates. The resulting deep network is not only generalizing the importance of temporal feature, but it also allows the interpretable feature representation and classification simultaneously. Detailed comparative case study on the benchmark Tennessee Eastman process demonstrates that the proposed method is competitive with the classical long-short term memory (LSTM) network for the fault detection and diagnosis.

Image Tampering Localization Using a Dense Fully Convolutional Network

The emergence of powerful image editing software has substantially facilitated digital image tampering, leading to many security issues. Hence, it is urgent to identify tampered images and localize tampered regions. Although much attention has been devoted to image tampering localization in recent years, it is still challenging to perform tampering localization in practical forensic applications. The reasons include the difficulty of learning discriminative representations of tampering traces and the lack of realistic tampered images for training. Since Photoshop is widely used for image tampering in practice, this paper attempts to address the issue of tampering localization by focusing on the detection of commonly used editing tools and operations in Photoshop. In order to well capture tampering traces, a fully convolutional encoder-decoder architecture is designed, where dense connections and dilated convolutions are adopted for achieving better localization performance. In order to effectively train a model in the case of insufficient tampered images, we design a training data generation strategy by resorting to Photoshop scripting, which can imitate human manipulations and generate large-scale training samples. Extensive experimental results show that the proposed approach outperforms state-of-the-art competitors when the model is trained with only generated images or fine-tuned with a small amount of realistic tampered images. The proposed method also has good robustness against some common post-processing operations.

A Novel Spatio-Temporal 3D Convolutional Encoder-Decoder Network for Dynamic Saliency Prediction

As human beings are living in an always changing environment, predicting saliency maps from dynamic visual stimulus is of importance for modeling human visual system. Compared with human behavior, recent models based on LSTM and 3DCNN are still not good enough due to the limitation in spatio-temporal feature representation. In this paper, a novel 3D convolutional encoder-decoder architecture is proposed for saliency prediction on dynamic scenes. The encoder consists of two subnetworks to extract both spatial and temporal features in parallel with intermediate fusion, respectively. The saliency map is produced in decoder by firstly enlarging features in spatial dimensions and then aggregating temporal information. Specially designed structures can transfer pooling indices from encoder to decoder, which helps the generation of location-aware saliency maps. The proposed network can be trained and inferred in an end-to-end manner. Experimental results on benchmark DHF1K show that the proposed model achieves the state-of-the-art performance on key metrics including both normalized scanpath saliency and Pearson's correlation coefficient.

DiSegNet: A deep dilated convolutional encoder-decoder architecture for lymph node segmentation on PET/CT images.

Automated lymph node (LN) recognition and segmentation from cross-sectional medical images is an important step for the automated diagnostic assessment of patients with cancer. Yet, it is still a difficult task owing to the low contrast of LNs and surrounding soft tissues as well as due to the variation in nodal size and shape. In this paper, we present a novel LN segmentation method based on a newly designed neural network for positron emission tomography/computed tomography (PET/CT) images. This work communicates two problems involved in LN segmentation task. Firstly, an efficient loss function named cosine-sine (CS) is proposed for the voxel class imbalance problem in the convolution network training process. Second, a multi-stage and multi-scale Atrous (Dilated) spatial pyramid pooling sub-module, named MS-ASPP, is introduced to the encoder-decoder architecture (SegNet), which aims to make use of multi-scale information to improve the performance of LN segmentation. The new architecture is named DiSegNet (Dilated SegNet). Four-fold cross-validation is performed on 63 PET/CT data sets. In each experiment, 10 data sets are selected randomly for testing and the other 53 for training. The results show that we reach an average 77 % Dice similarity coefficient score with CS loss function by trained DiSegNet, compared to a baseline method SegNet by cross-entropy (CE) with 71 % Dice similarity coefficient. The performance of the proposed DiSegNet with CS loss function suggests its potential clinical value for disease quantification.

Deep Convolutional Encoder-Decoder algorithm for MRI brain reconstruction.

Compressed Sensing Magnetic Resonance Imaging (CS-MRI) could be considered a challenged task since it could be designed as an efficient technique for fast MRI acquisition which could be highly beneficial for several clinical routines. In fact, it could grant better scan quality by reducing motion artifacts amount as well as the contrast washout effect. It offers also the possibility to reduce the exploration cost and the patient's anxiety. Recently, Deep Learning Neuronal Network (DL) has been suggested in order to reconstruct MRI scans with conserving the structural details and improving parallel imaging-based fast MRI. In this paper, we propose Deep Convolutional Encoder-Decoder architecture for CS-MRI reconstruction. Such architecture bridges the gap between the non-learning techniques, using data from only one image, and approaches using large training data. The proposed approach is based on autoencoder architecture divided into two parts: an encoder and a decoder. The encoder as well as the decoder has essentially three convolutional blocks. The proposed architecture has been evaluated through two databases: Hammersmith dataset (for the normal scans) and MICCAI 2018 (for pathological MRI). Moreover, we extend our model to cope with noisy pathological MRI scans. The normalized mean square error (NMSE), the peak-to-noise ratio (PSNR), and the structural similarity index (SSIM) have been adopted as evaluation metrics in order to evaluate the proposed architecture performance and to make a comparative study with the state-of-the-art reconstruction algorithms. The higher PSNR and SSIM values as well as the lowest NMSE values could attest that the proposed architecture offers better reconstruction and preserves textural image details. Furthermore, the running time is about 0.8s, which is suitable for real-time processing. Such results could encourage the neurologist to adopt it in their clinical routines. Graphical abstract.

Gated dynamic convolutions with deep layer fusion for abstractive document summarization

We present a novel abstractive document summarization based on the recently proposed dynamic convolutional encoder-decoder architectures. We address several aspects of summarization that are not well modeled by the basic architecture, by integrating multiple layers of the encoder, controlling information flow in the hierarchy, and exploiting external knowledge. First, we propose a simple and efficient deep layer fusion to extract salient information from the encoder layers. Second, we propose a gating mechanism to control and maintain important contextual information through the encoder-decoder layers into dynamic convolutions. Lastly, we put part-of-speech information into the model as external knowledge to better predict filters for dynamic convolutions. We evaluate our model using ROUGE metrics on three different datasets: CNN-DM, NEWSROOM-ABS, and XSUM. Experimental results show that the proposed model outperforms the state-of-the-art abstractive models on NEWSROOM-ABS and XSUM and shows comparable scores on CNN-DM.

Automatic extraction of cell nuclei using dilated convolutional network

Pathological examination has been done manually by visual inspection of hematoxylin and eosin (H&E)-stained images. However, this process is labor intensive, prone to large variations, and lacking reproducibility in the diagnosis of a tumor. We aim to develop an automatic workflow to extract different cell nuclei found in cancerous tumors portrayed in digital renderings of the H&E-stained images. For a given image, we propose a semantic pixel-wise segmentation technique using dilated convolutions. The architecture of our dilated convolutional network (DCN) is based on SegNet, a deep convolutional encoder-decoder architecture. For the encoder, all the max pooling layers in the SegNet are removed and the convolutional layers are replaced by dilated convolution layers with increased dilation factors to preserve image resolution. For the decoder, all max unpooling layers are removed and the convolutional layers are replaced by dilated convolution layers with decreased dilation factors to remove gridding artifacts. We show that dilated convolutions are superior in extracting information from textured images. We test our DCN network on both synthetic data sets and a public available data set of H&E-stained images and achieve better results than the state of the art.

Building extraction from VHR remote sensing imagery by combining an improved deep convolutional encoder-decoder architecture and historical land use vector map

ABSTRACT Building extraction has attracted considerable attention in the field of remote sensing image analysis. Fully convolutional network modelling is a recently developed technique that is capable of significantly enhancing building extraction accuracy. It is a prominent branch of deep learning and uses advanced state-of-the-art techniques, especially with regard to building segmentation. In this paper, we present an enhanced deep convolutional encoder-decoder (DCED) network by incorporating historical land use vector maps (HVMs) customized for building extraction. The approach combines enhanced DCED architecture with multi-scale image pyramid for pixel-wise building segmentation. The improved DCED network, together with symmetrical dense-shortcut connection structures, is employed to establish the encoders for automatic extraction of building features. The feature maps from early layers were fused with more discriminative feature maps from the deeper layers through ‘Res path’ skip connections for superior building extraction accuracy. To further reduce the occurrence of falsely segmented buildings, and to sharpen the buildings’ boundaries, the new temporal testing image is segmented under the constraints of an HVM. A majority voting strategy is employed to ensure the homogeneity of the building objects as the post-processing method. Experimental results indicate that the proposed approach exhibits competitive quantitative and qualitative performance, effectively alleviating the salt-and-pepper phenomenon and block effects, and retaining the edge structures of buildings. Compared with other state-of-the-art methods, our method demonstrably achieves the optimal final accuracies.

Healthy versus pathological learning transferability in shoulder muscle MRI segmentation using deep convolutional encoder-decoders

Fully-automated segmentation of pathological shoulder muscles in patients with musculo-skeletal diseases is a challenging task due to the huge variability in muscle shape, size, location, texture and injury. A reliable automatic segmentation method from magnetic resonance images could greatly help clinicians to diagnose pathologies, plan therapeutic interventions and predict interventional outcomes while eliminating time consuming manual segmentation. The purpose of this work is three-fold. First, we investigate the feasibility of automatic pathological shoulder muscle segmentation using deep learning techniques, given a very limited amount of available annotated pediatric data. Second, we address the learning transferability from healthy to pathological data by comparing different learning schemes in terms of model generalizability. Third, extended versions of deep convolutional encoder-decoder architectures using encoders pre-trained on non-medical data are proposed to improve the segmentation accuracy. Methodological aspects are evaluated in a leave-one-out fashion on a dataset of 24 shoulder examinations from patients with unilateral obstetrical brachial plexus palsy and focus on 4 rotator cuff muscles (deltoid, infraspinatus, supraspinatus and subscapularis). The most accurate segmentation model is partially pre-trained on the large-scale ImageNet dataset and jointly exploits inter-patient healthy and pathological annotated data. Its performance reaches Dice scores of 82.4%, 82.0%, 71.0% and 82.8% for deltoid, infraspinatus, supraspinatus and subscapularis muscles. Absolute surface estimation errors are all below 83 mm2 except for supraspinatus with 134.6 mm2. The contributions of our work offer new avenues for inferring force from muscle volume in the context of musculo-skeletal disorder management.