Publications

Deep learning for detection and segmentation of artefact and disease instances in gastrointestinal endoscopy

Abstract

The Endoscopy Computer Vision Challenge (EndoCV) is a crowd-sourcing initiative to address eminent problems in developing reliable computer aided detection and diagnosis endoscopy systems and suggest a pathway for clinical translation of technologies. Whilst endoscopy is a widely used diagnostic and treatment tool for hollow-organs, there are several core challenges often faced by endoscopists, mainly: 1) presence of multi-class artefacts that hinder their visual interpretation, and 2) difficulty in identifying subtle precancerous precursors and cancer abnormalities. Artefacts often affect the robustness of deep learning methods applied to the gastrointestinal tract organs as they can be confused with tissue of interest. EndoCV2020 challenges are designed to address research questions in these remits. In this paper, we present a summary of methods developed by the top 17 teams and provide an objective comparison of state-of-the-art methods and methods designed by the participants for two sub-challenges: i) artefact detection and segmentation (EAD2020), and ii) disease detection and segmentation (EDD2020). Multi-center, multi-organ, multi-class, and multi-modal clinical endoscopy datasets were compiled for both EAD2020 and EDD2020 sub-challenges. The out-of-sample generalization ability of detection algorithms was also evaluated. Whilst most teams focused on accuracy improvements, only a few methods hold credibility for clinical usability. The best performing teams provided solutions to tackle class imbalance, and variabilities in size, origin, modality and occurrences by exploring data augmentation, data fusion, and optimal class thresholding techniques.

Continue reading

Combining deep learning and mathematical morphology for historical map segmentation

Abstract

The digitization of historical maps enables the study of ancient, fragile, unique, and hardly accessible information sources. Main map features can be retrieved and tracked through the time for subsequent thematic analysis. The goal of this work is the vectorization step, i.e., the extraction of vector shapes of the objects of interest from raster images of maps. We are particularly interested in closed shape detection such as buildings, building blocks, gardens, rivers, etc. in order to monitor their temporal evolution. Historical map images present significant pattern recognition challenges. The extraction of closed shapes by using traditional Mathematical Morphology (MM) is highly challenging due to the overlapping of multiple map features and texts. Moreover, state-of-the-art Convolutional Neural Networks (CNN) are perfectly designed for content image filtering but provide no guarantee about closed shape detection. Also, the lack of textural and color information of historical maps makes it hard for CNN to detect shapes that are represented by only their boundaries. Our contribution is a pipeline that combines the strengths of CNN (efficient edge detection and filtering) and MM (guaranteed extraction of closed shapes) in order to achieve such a task. The evaluation of our approach on a public dataset shows its effectiveness for extracting the closed boundaries of objects in historical maps.

Continue reading

Going beyond p-convolutions to learn grayscale morphological operators

Abstract

Integrating mathematical morphology operations within deep neural networks has been subject to increasing attention lately. However, replacing standard convolution layers with erosions or dilations is particularly challenging because the min and max operations are not differentiable. Relying on the asymptotic behavior of the counter-harmonic mean, p-convolutional layers were proposed as a possible workaround to this issue since they can perform pseudo-dilation or pseudo-erosion operations (depending on the value of their inner parameter p), and very promising results were reported. In this work, we present two new morphological layers based on the same principle as the p-convolutional layer while circumventing its principal drawbacks, and demonstrate their potential interest in further implementations within deep convolutional neural network architectures.

Continue reading

On some associations between mathematical morphology and artificial intelligence

Abstract

This paper aims at providing an overview of the use of mathematical morphology, in its algebraic setting, in several fields of artificial intelligence (AI). Three domains of AI will be covered. In the first domain, mathematical morphology operators will be expressed in some logics (propositional, modal, description logics) to answer typical questions in knowledge representation and reasoning, such as revision, fusion, explanatory relations, satisfying usual postulates. In the second domain, spatial reasoning will benefit from spatial relations modeled using fuzzy sets and morphological operators, with applications in model-based image understanding. In the third domain, interactions between mathematical morphology and deep learning will be detailed. Morphological neural networks were introduced as an alternative to classical architectures, yielding a new geometry in decision surfaces. Deep networks were also trained to learn morphological operators and pipelines, and morphological algorithms were used as companion tools to machine learning, for pre/post processing or even regularization purposes. These ideas have known a large resurgence in the last few years and new ones are emerging.

Continue reading

A machine learning based splitting heuristic for divide-and-conquer solvers

Abstract

In this paper, we present a machine learning based splitting heuristic for divide-and-conquer parallel Boolean SAT solvers. Splitting heuristics, whether they are look-ahead or look-back, are designed using proxy metrics, which when optimized, approximate the true metric of minimizing solver runtime on sub-formulas resulting from a split. The rationale for such metrics is that they have been empirically shown to be excellent proxies for runtime of solvers, in addition to being cheap to compute in an online fashion. However, the design of traditional splitting methods are often ad-hoc and do not leverage the copious amounts of data that solvers generate. To address the above-mentioned issues, we propose a machine learning based splitting heuristic that leverages the features of input formulas and data generated during the run of a divide-and-conquer (DC) parallel solver. More precisely, we reformulate the splitting problem as a ranking problem and develop two machine learning models for pairwise ranking and computing the minimum ranked variable. Our model can compare variables according to their splitting quality, which is based on a set of features extracted from structural properties of the input formula, as well as dynamic probing statistics, collected during the solver’s run. We derive the true labels through offline collection of runtimes of a parallel DC solver on sample formulas and variables within them. At each splitting point, we generate a predicted ranking (pairwise or minimum rank) of candidate variables and split the formula on the top variable. We implemented our heuristic in the Painless parallel SAT framework and evaluated our solver on a set of cryptographic instances encoding the SHA-1 preimage as well as SAT competition 2018 and 2019 benchmarks. We solve significantly more instances compared to the baseline Painless solver and outperform top divide-and-conquer solvers from recent SAT competitions, such as Treengeling. Furthermore, we are much faster than these top solvers on cryptographic benchmarks.

Continue reading

State-of-the-art speaker recognition with neural network embeddings in NIST SRE18 and speakers in the wild evaluations

Abstract

Continue reading

Stacked and parallel U-nets with multi-output for myocardial pathology segmentation

Abstract

In the field of medical imaging, many different image modalities contain different information, helping practitionners to make diagnostic, follow-up, etc. To better analyze images, mixing multi-modalities information has become a trend. This paper provides one cascaded UNet framework and uses three different modalities (the late gadolinium enhancement (LGE) CMR sequence, the balanced- Steady State Free Precession (bSSFP) cine sequence and the T2-weighted CMR) to complete the segmentation of the myocardium, scar and edema in the context of the MICCAI 2020 myocardial pathology segmentation combining multi-sequence CMR Challenge dataset (MyoPS 2020). We evaluate the proposed method with 5-fold-cross-validation on the MyoPS 2020 dataset.

Continue reading

A global benchmark of algorithms for segmenting the left atrium from late gadolinium-enhanced cardiac magnetic resonance imaging

Abstract

Segmentation of medical images, particularly late gadolinium-enhanced magnetic resonance imaging (LGE-MRI) used for visualizing diseased atrial structures, is a crucial first step for ablation treatment of atrial fibrillation. However, direct segmentation of LGE-MRIs is challenging due to the varying intensities caused by contrast agents. Since most clinical studies have relied on manual, labor-intensive approaches, automatic methods are of high interest, particularly optimized machine learning approaches. To address this, we organized the 2018 Left Atrium Segmentation Challenge using 154 3D LGE-MRIs, currently the world’s largest atrial LGE-MRI dataset, and associated labels of the left atrium segmented by three medical experts, ultimately attracting the participation of 27 international teams. In this paper, extensive analysis of the submitted algorithms using technical and biological metrics was performed by undergoing subgroup analysis and conducting hyper-parameter analysis, offering an overall picture of the major design choices of convolutional neural networks (CNNs) and practical considerations for achieving state-of-the-art left atrium segmentation. Results show that the top method achieved a Dice score of 93.2% and a mean surface to surface distance of 0.7 mm, significantly outperforming prior state-of-the-art. Particularly, our analysis demonstrated that double sequentially used CNNs, in which a first CNN is used for automatic region-of-interest localization and a subsequent CNN is used for refined regional segmentation, achieved superior results than traditional methods and machine learning approaches containing single CNNs. This large-scale benchmarking study makes a significant step towards much-improved segmentation methods for atrial LGE-MRIs, and will serve as an important benchmark for evaluating and comparing the future works in the field. Furthermore, the findings from this study can potentially be extended to other imaging datasets and modalities, having an impact on the wider medical imaging community.

Continue reading

PAIP 2019: Liver cancer segmentation challenge

Abstract

Pathology Artificial Intelligence Platform (PAIP) is a free research platform in support of pathological artificial intelligence (AI). The main goal of the platform is to construct a high-quality pathology learning data set that will allow greater accessibility. The PAIP Liver Cancer Segmentation Challenge, organized in conjunction with the Medical Image Computing and Computer Assisted Intervention Society (MICCAI 2019), is the first image analysis challenge to apply PAIP datasets. The goal of the challenge was to evaluate new and existing algorithms for automated detection of liver cancer in whole-slide images (WSIs). Additionally, the PAIP of this year attempted to address potential future problems of AI applicability in clinical settings. In the challenge, participants were asked to use analytical data and statistical metrics to evaluate the performance of automated algorithms in two different tasks. The participants were given the two different tasks: Task 1 involved investigating Liver Cancer Segmentation and Task 2 involved investigating Viable Tumor Burden Estimation. There was a strong correlation between high performance of teams on both tasks, in which teams that performed well on Task 1 also performed well on Task 2. After evaluation, we summarized the top 11 team’s algorithms. We then gave pathological implications on the easily predicted images for cancer segmentation and the challenging images for viable tumor burden estimation. Out of the 231 participants of the PAIP challenge datasets, a total of 64 were submitted from 28 team participants. The submitted algorithms predicted the automatic segmentation on the liver cancer with WSIs to an accuracy of a score estimation of 0.78. The PAIP challenge was created in an effort to combat the lack of research that has been done to address Liver cancer using digital pathology. It remains unclear of how the applicability of AI algorithms created during the challenge can affect clinical diagnoses. However, the results of this dataset and evaluation metric provided has the potential to aid the development and benchmarking of cancer diagnosis and segmentation.

Continue reading

Do not treat boundaries and regions differently: An example on heart left atrial segmentation

Abstract

Atrial fibrillation is the most common heart rhythm disease. Due to a lack of understanding in matter of underlying atrial structures, current treatments are still not satisfying. Recently, with the popularity of deep learning, many segmentation methods based on fully convolutional networks have been proposed to analyze atrial structures, especially from late gadolinium-enhanced magnetic resonance imaging. However, two problems still occur: 1) segmentation results include the atrial- like background; 2) boundaries are very hard to segment. Most segmentation approaches design a specific network that mainly focuses on the regions, to the detriment of the boundaries. Therefore, this paper proposes an attention full convolutional network framework based on the ResNet-101 architecture, which focuses on boundaries as much as on regions. The additional attention module is added to have the network pay more attention on regions and then to reduce the impact of the misleading similarity of neighboring tissues. We also use a hybrid loss composed of a region loss and a boundary loss to treat boundaries and regions at the same time. We demonstrate the efficiency of the proposed approach on the MICCAI 2018 Atrial Segmentation Challenge public dataset.

Continue reading