CT-guided spatial normalization of nuclear hybrid imaging adapted to enlarged ventricles: Impact on striatal uptake quantification.

INTRODUCTION: Spatial normalization is a prerequisite step for the quantitative analysis of SPECT or PET brain images using volume-of-interest (VOI) template or voxel-based analysis. MRI-guided spatial normalization is the gold standard, but the wide use of PET/CT or SPECT/CT in routine clinical practice makes CT-guided spatial normalization a necessary alternative. Ventricular enlargement is observed with aging, and it hampers the spatial normalization of the lateral ventricles and striatal regions, limiting their analysis. The aim of the present study was to propose a robust spatial normalization method based on CT scans that takes into account features of the aging brain to reduce bias in the CT-guided striatal analysis of SPECT images. METHODS: We propose an enhanced CT-guided spatial normalization pipeline based on SPM12. Performance of the proposed pipeline was assessed on visually normal [123I]-FP-CIT SPECT/CT images. SPM12 default CT-guided spatial normalization was used as reference method. The metrics assessed were the overlap between the spatially normalized lateral ventricles and caudate/putamen VOIs, and the computation of caudate and putamen specific binding ratios (SBR). RESULTS: In total 231 subjects (mean age ± SD = 61.9 ± 15.5 years) were included in the statistical analysis. The mean overlap between the spatially normalized lateral ventricles of subjects and the caudate VOI and the mean SBR of caudate were respectively 38.40 % (± SD = 19.48 %) of the VOI and 1.77 (± 0.79) when performing SPM12 default spatial normalization. The mean overlap decreased to 9.13 % (± SD = 1.41 %, P < 0.001) of the VOI and the SBR of caudate increased to 2.38 (± 0.51, P < 0.0001) when performing the proposed pipeline. Spatially normalized lateral ventricles did not overlap with putamen VOI using either method. The mean putamen SBR value derived from the proposed spatial normalization (2.75 ± 0.54) was not significantly different from that derived from the default SPM12 spatial normalization (2.83 ± 0.52, P > 0.05). CONCLUSION: The automatic CT-guided spatial normalization used herein led to a less biased spatial normalization of SPECT images, hence an improved semi-quantitative analysis. The proposed pipeline could be implemented in clinical routine to perform a more robust SBR computation using hybrid imaging.

A Weakly Supervised Gradient Attribution Constraint for Interpretable Classification and Anomaly Detection.

The lack of interpretability of deep learning reduces understanding of what happens when a network does not work as expected and hinders its use in critical fields like medicine, which require transparency of decisions. For example, a healthy vs pathological classification model should rely on radiological signs and not on some training dataset biases. Several post-hoc models have been proposed to explain the decision of a trained network. However, they are very seldom used to enforce interpretability during training and none in accordance with the classification. In this paper, we propose a new weakly supervised method for both interpretable healthy vs pathological classification and anomaly detection. A new loss function is added to a standard classification model to constrain each voxel of healthy images to drive the network decision towards the healthy class according to gradient-based attributions. This constraint reveals pathological structures for patient images, allowing their unsupervised segmentation. Moreover, we advocate both theoretically and experimentally, that constrained training with the simple Gradient attribution is similar to constraints with the heavier Expected Gradient, consequently reducing the computational cost. We also propose a combination of attributions during the constrained training making the model robust to the attribution choice at inference. Our proposition was evaluated on two brain pathologies: tumors and multiple sclerosis. This new constraint provides a more relevant classification, with a more pathology-driven decision. For anomaly detection, the proposed method outperforms state-of-the-art especially on difficult multiple sclerosis lesions segmentation task with a 15 points Dice improvement.

Development of normal pressure hydrocephalus following post-traumatic external hydrocephalus in an adult patient.

We report the case of 74-year-old patient suspected of post-traumatic external hydrocephalus (EH) following a mild traumatic brain injury with a progressive neurological decline and a concomitant enlargement of subarachnoid spaces without ventriculomegaly on CT scan. A lumbar puncture revealed raised ICP and a careful CSF withdrawal was performed, resulting in an immediate neurological improvement, confirming the diagnosis of EH. During the 20-month follow-up, the patient presented progressive signs of normal pressure hydrocephalus (NPH): gait and cognitive decline, ventriculomegaly and the lumbar infusion study confirmed disturbed CSF dynamics. The patient underwent a ventriculoperitoneal shunt surgery, resulting in a long-lasting improvement.

Multiple sclerosis clinical forms classification with graph convolutional networks based on brain morphological connectivity.

Multiple Sclerosis (MS) is an autoimmune disease that combines chronic inflammatory and neurodegenerative processes underlying different clinical forms of evolution, such as relapsing-remitting, secondary progressive, or primary progressive MS. This identification is usually performed by clinical evaluation at the diagnosis or during the course of the disease for the secondary progressive phase. In parallel, magnetic resonance imaging (MRI) analysis is a mandatory diagnostic complement. Identifying the clinical form from MR images is therefore a helpful and challenging task. Here, we propose a new approach for the automatic classification of MS forms based on conventional MRI (i.e., T1-weighted images) that are commonly used in clinical context. For this purpose, we investigated the morphological connectome features using graph based convolutional neural network. Our results obtained from the longitudinal study of 91 MS patients highlight the performance (F1-score) of this approach that is better than state-of-the-art as 3D convolutional neural networks. These results open the way for clinical applications such as disability correlation only using T1-weighted images.

PET image enhancement using artificial intelligence for better characterization of epilepsy lesions.

Introduction: [18F]fluorodeoxyglucose ([18F]FDG) brain PET is used clinically to detect small areas of decreased uptake associated with epileptogenic lesions, e.g., Focal Cortical Dysplasias (FCD) but its performance is limited due to spatial resolution and low contrast. We aimed to develop a deep learning-based PET image enhancement method using simulated PET to improve lesion visualization. Methods: We created 210 numerical brain phantoms (MRI segmented into 9 regions) and assigned 10 different plausible activity values (e.g., GM/WM ratios) resulting in 2100 ground truth high quality (GT-HQ) PET phantoms. With a validated Monte-Carlo PET simulator, we then created 2100 simulated standard quality (S-SQ) [18F]FDG scans. We trained a ResNet on 80% of this dataset (10% used for validation) to learn the mapping between S-SQ and GT-HQ PET, outputting a predicted HQ (P-HQ) PET. For the remaining 10%, we assessed Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Root Mean Squared Error (RMSE) against GT-HQ PET. For GM and WM, we computed recovery coefficients (RC) and coefficient of variation (COV). We also created lesioned GT-HQ phantoms, S-SQ PET and P-HQ PET with simulated small hypometabolic lesions characteristic of FCDs. We evaluated lesion detectability on S-SQ and P-HQ PET both visually and measuring the Relative Lesion Activity (RLA, measured activity in the reduced-activity ROI over the standard-activity ROI). Lastly, we applied our previously trained ResNet on 10 clinical epilepsy PETs to predict the corresponding HQ-PET and assessed image quality and confidence metrics. Results: Compared to S-SQ PET, P-HQ PET improved PNSR, SSIM and RMSE; significatively improved GM RCs (from 0.29 ± 0.03 to 0.79 ± 0.04) and WM RCs (from 0.49 ± 0.03 to 1 ± 0.05); mean COVs were not statistically different. Visual lesion detection improved from 38 to 75%, with average RLA decreasing from 0.83 ± 0.08 to 0.67 ± 0.14. Visual quality of P-HQ clinical PET improved as well as reader confidence. Conclusion: P-HQ PET showed improved image quality compared to S-SQ PET across several objective quantitative metrics and increased detectability of simulated lesions. In addition, the model generalized to clinical data. Further evaluation is required to study generalization of our method and to assess clinical performance in larger cohorts.

Deep learning in veterinary medicine, an approach based on CNN to detect pulmonary abnormalities from lateral thoracic radiographs in cats.

Thoracic radiograph (TR) is a complementary exam widely used in small animal medicine which requires a sharp analysis to take full advantage of Radiographic Pulmonary Pattern (RPP). Although promising advances have been made in deep learning for veterinary imaging, the development of a Convolutional Neural Networks (CNN) to detect specifically RPP from feline TR images has not been investigated. Here, a CNN based on ResNet50V2 and pre-trained on ImageNet is first fine-tuned on human Chest X-rays and then fine-tuned again on 500 annotated TR images from the veterinary campus of VetAgro Sup (Lyon, France). The impact of manual segmentation of TR's intrathoracic area and enhancing contrast method on the CNN's performances has been compared. To improve classification performances, 200 networks were trained on random shuffles of training set and validation set. A voting approach over these 200 networks trained on segmented TR images produced the best classification performances and achieved mean Accuracy, F1-Score, Specificity, Positive Predictive Value and Sensitivity of 82%, 85%, 75%, 81% and 88% respectively on the test set. Finally, the classification schemes were discussed in the light of an ensemble method of class activation maps and confirmed that the proposed approach is helpful for veterinarians.

Deep learning detection of nanoparticles and multiple object tracking of their dynamic evolution during in situ ETEM studies.

In situ transmission electron microscopy (TEM) studies of dynamic events produce large quantities of data especially under the form of images. In the important case of heterogeneous catalysis, environmental TEM (ETEM) under gas and temperature allows to follow a large population of supported nanoparticles (NPs) evolving under reactive conditions. Interpreting properly large image sequences gives precious information on the catalytic properties of the active phase by identifying causes for its deactivation. To perform a quantitative, objective and robust treatment, we propose an automatic procedure to track nanoparticles observed in Scanning ETEM (STEM in ETEM). Our approach involves deep learning and computer vision developments in multiple object tracking. At first, a registration step corrects the image displacements and misalignment inherent to the in situ acquisition. Then, a deep learning approach detects the nanoparticles on all frames of video sequences. Finally, an iterative tracking algorithm reconstructs their trajectories. This treatment allows to deduce quantitative and statistical features about their evolution or motion, such as a Brownian behavior and merging or crossing events. We treat the case of in situ calcination of palladium (oxide) / delta-alumina, where the present approach allows a discussion of operating processes such as Ostwald ripening or NP aggregative coalescence.

Extracellular AGR2 triggers lung tumour cell proliferation through repression of p21CIP1.

The human Anterior GRadient 2 (AGR2) protein is an Endoplasmic Reticulum (ER)-resident protein which belongs to the Protein-Disulfide Isomerase (PDI) superfamily and is involved to productive protein folding in the ER. As such AGR2, often found overexpressed in adenocarcinomas, contributes to tumour development by enhancing ER proteostasis. We previously demonstrated that AGR2 is secreted (extracellular AGR2 (eAGR2)) in the tumour microenvironment and plays extracellular roles independent of its ER functions. Herein, we show that eAGR2 triggers cell proliferation and characterize the underlying molecular mechanisms. We demonstrate that eAGR2 enhances tumour cell growth by repressing the tumour suppressor p21CIP1. Our findings shed light on a novel mechanism through which eAGR2 behaves as a growth factor in the tumour microenvironment, independently of its ER function, thus promoting tumour cell growth through repression of p21CIP1. Our results provide a rationale for targeting eAGR2/p21CIP1-based signalling as a potential therapeutic target to impede tumour growth.

Quantitative Magnetic Resonance Imaging Assessment of the Quadriceps Changes during an Extreme Mountain Ultramarathon.

INTRODUCTION/PURPOSE: Extreme ultra-endurance races are growing in popularity, but their effects on skeletal muscles remain mostly unexplored. This longitudinal study explores physiological changes in mountain ultramarathon athletes' quadriceps using quantitative magnetic resonance imaging (MRI) coupled with serological biomarkers. The study aimed to monitor the longitudinal effect of the race and recovery and to identify local inflammatory and metabolic muscle responses by codetection of biological markers. METHODS: An automatic image processing framework was designed to extract imaging-based biomarkers from quantitative MRI acquisitions of the upper legs of 20 finishers at three time points. The longitudinal effect of the race was demonstrated by analyzing the image markers with dedicated biostatistical analysis. RESULTS: Our framework allows for a reliable calculation of statistical data not only inside the whole quadriceps volume but also within each individual muscle head. Local changes in MRI parameters extracted from quantitative maps were described and found to be significantly correlated with principal serological biomarkers of interest. A decrease in the PDFF after the race and a stable paramagnetic susceptibility value were found. Pairwise post hoc tests suggested that the recovery process differs among the muscle heads. CONCLUSIONS: This longitudinal study conducted during a prolonged and extreme mechanical stress showed that quantitative MRI-based markers of inflammation and metabolic response can detect local changes related to the prolonged exercise, with differentiated involvement of each head of the quadriceps muscle as expected in such eccentric load. Consistent and efficient extraction of the local biomarkers enables to highlight the interplay/interactions between blood and MRI biomarkers. This work indeed proposes an automatized analytic framework to tackle the time-consuming and mentally exhausting segmentation task of muscle heads in large multi-time-point cohorts.

Patients Lung Derived Tumoroids (PLDTs) to model therapeutic response.

Preclinical lung cancer models are essential for a basic understanding of lung cancer biology and its translation into efficient treatment options for affected patients. Lung cancer cell lines and xenografts derived directly from human lung tumors have proven highly valuable in fundamental oncology research and anticancer drug discovery. Both models inherently comprise advantages and caveats that have to be accounted for. Recently, we have enabled reliable in vitro culture techniques from lung cancer biopsies as Patients Lung Derived Tumoroids (PLDTs). This breakthrough provides the possibility of high-throughput drug screening covering the spectrum of lung cancer phenotypes seen clinically. We have adapted and optimized our in vitro three-dimensional model as a preclinical lung cancer model to recapitulate the tumor microenvironment (TME) using matrix reconstitution. Hence, we developed directly PLDTs to screen for chemotherapeutics and radiation treatment. This original model will enable precision medicine to become a reality, allowing a given patient sample to be screened for effective ex vivo therapeutics, aiming at tailoring of treatments specific to that individual. Hence, this tool can enhance clinical outcomes and avoid morbidity due to ineffective therapies.

LU-Net: A Multistage Attention Network to Improve the Robustness of Segmentation of Left Ventricular Structures in 2-D Echocardiography.

Segmentation of cardiac structures is one of the fundamental steps to estimate volumetric indices of the heart. This step is still performed semiautomatically in clinical routine and is, thus, prone to interobserver and intraobserver variabilities. Recent studies have shown that deep learning has the potential to perform fully automatic segmentation. However, the current best solutions still suffer from a lack of robustness in terms of accuracy and number of outliers. The goal of this work is to introduce a novel network designed to improve the overall segmentation accuracy of left ventricular structures (endocardial and epicardial borders) while enhancing the estimation of the corresponding clinical indices and reducing the number of outliers. This network is based on a multistage framework where both the localization and segmentation steps are optimized jointly through an end-to-end scheme. Results obtained on a large open access data set show that our method outperforms the current best-performing deep learning solution with a lighter architecture and achieved an overall segmentation accuracy lower than the intraobserver variability for the epicardial border (i.e., on average a mean absolute error of 1.5 mm and a Hausdorff distance of 5.1mm) with 11% of outliers. Moreover, we demonstrate that our method can closely reproduce the expert analysis for the end-diastolic and end-systolic left ventricular volumes, with a mean correlation of 0.96 and a mean absolute error of 7.6 ml. Concerning the ejection fraction of the left ventricle, results are more contrasted with a mean correlation coefficient of 0.83 and an absolute mean error of 5.0%, producing scores that are slightly below the intraobserver margin. Based on this observation, areas for improvement are suggested.

A Pilot Study on Convolutional Neural Networks for Motion Estimation From Ultrasound Images.

In recent years, deep learning (DL) has been successfully applied to the analysis and processing of ultrasound images. To date, most of this research has focused on segmentation and view recognition. This article benchmarks different convolutional neural network algorithms for motion estimation in ultrasound imaging. We evaluated and compared several networks derived from FlowNet2, one of the most efficient architectures in computer vision. The networks were tested with and without transfer learning, and the best configuration was compared against the particle imaging velocimetry method, a popular state-of-the-art block-matching algorithm. Rotations are known to be difficult to track from ultrasound images due to a significant speckle decorrelation. We thus focused on the images of rotating disks, which could be tracked through speckle features only. Our database consisted of synthetic and in vitro B-mode images after log compression and covered a large range of rotational speeds. One of the FlowNet2 subnetworks, FlowNet2SD, produced competitive results with a motion field error smaller than 1 pixel on real data after transfer learning based on the simulated data. These errors remain small for a large velocity range without the need for hyperparameter tuning, which indicates the high potential and adaptability of DL solutions to motion estimation in ultrasound imaging.

Automatic myocardial ischemic lesion detection on magnetic resonance perfusion weighted imaging prior perfusion quantification: A pre-modeling strategy.

Even if cardiovascular magnetic resonance (CMR) perfusion imaging has proven its relevance for visual detection of ischemia, myocardial blood flow (MBF) quantification at the voxel observation scale remains challenging. Integration of an automated segmentation step, prior to perfusion index estimation, might be a significant reconstruction component that could allow sustainable assumptions and constraint enlargement prior to advanced modeling. Current clustering techniques, such as bullseye representation or manual delineation, are not designed to discriminate voxels belonging to the lesion from healthy areas. Hence, the resulting average time-intensity curve, which is assumed to represent the dynamic contrast enhancement inside of a lesion, might be contaminated by voxels with perfectly healthy microcirculation. This study introduces a hierarchical lesion segmentation approach based on time-intensity curve features that considers the spatial particularities of CMR myocardial perfusion. A first k-means clustering approach enables this method to perform coarse clustering, which is refined by a novel spatiotemporal region-growing (STRG) segmentation, thus ensuring spatial and time-intensity curve homogeneity. Over a cohort of 30 patients, myocardial blood flow (MBF) measured in voxels of lesion regions detected with STRG was significantly lower than in regions drawn manually (mean difference = 0.14, 95% CI [0.07, 0.2]) and defined with the bullseye template (mean difference = 0.25, 95% CI [0.17, 0.36]). Over the 90 analyzed slices, the median Dice score calculated against the ground truth ranged between 0.62 and 0.67, the inclusion coefficients ranged between 0.62 and 0.76 and the centroid distances ranged between 0.97 and 3.88 mm. Therefore, though these metrics highlight spatial differences, they could not be used as an index to evaluate the accuracy and performance of the method, which can only be attested by the variability of the MBF clinical index.

Deep Learning for Segmentation Using an Open Large-Scale Dataset in 2D Echocardiography.

Delineation of the cardiac structures from 2D echocardiographic images is a common clinical task to establish a diagnosis. Over the past decades, the automation of this task has been the subject of intense research. In this paper, we evaluate how far the state-of-the-art encoder-decoder deep convolutional neural network methods can go at assessing 2D echocardiographic images, i.e., segmenting cardiac structures and estimating clinical indices, on a dataset, especially, designed to answer this objective. We, therefore, introduce the cardiac acquisitions for multi-structure ultrasound segmentation dataset, the largest publicly-available and fully-annotated dataset for the purpose of echocardiographic assessment. The dataset contains two and four-chamber acquisitions from 500 patients with reference measurements from one cardiologist on the full dataset and from three cardiologists on a fold of 50 patients. Results show that encoder-decoder-based architectures outperform state-of-the-art non-deep learning methods and faithfully reproduce the expert analysis for the end-diastolic and end-systolic left ventricular volumes, with a mean correlation of 0.95 and an absolute mean error of 9.5 ml. Concerning the ejection fraction of the left ventricle, results are more contrasted with a mean correlation coefficient of 0.80 and an absolute mean error of 5.6%. Although these results are below the inter-observer scores, they remain slightly worse than the intra-observer's ones. Based on this observation, areas for improvement are defined, which open the door for accurate and fully-automatic analysis of 2D echocardiographic images.

Evaluation of noise and blur effects with SIRT-FISTA-TV reconstruction algorithm: Application to fast environmental transmission electron tomography.

Fast tomography in Environmental Transmission Electron Microscopy (ETEM) is of a great interest for in situ experiments where it allows to observe 3D real-time evolution of nanomaterials under operating conditions. In this context, we are working on speeding up the acquisition step to a few seconds mainly with applications on nanocatalysts. In order to accomplish such rapid acquisitions of the required tilt series of projections, a modern 4K high-speed camera is used, that can capture up to 100 images per second in a 2K binning mode. However, due to the fast rotation of the sample during the tilt procedure, noise and blur effects may occur in many projections which in turn would lead to poor quality reconstructions. Blurred projections make classical reconstruction algorithms inappropriate and require the use of prior information. In this work, a regularized algebraic reconstruction algorithm named SIRT-FISTA-TV is proposed. The performance of this algorithm using blurred data is studied by means of a numerical blur introduced into simulated images series to mimic possible mechanical instabilities/drifts during fast acquisitions. We also present reconstruction results from noisy data to show the robustness of the algorithm to noise. Finally, we show reconstructions with experimental datasets and we demonstrate the interest of fast tomography with an ultra-fast acquisition performed under environmental conditions, i.e. gas and temperature, in the ETEM. Compared to classically used SIRT and SART approaches, our proposed SIRT-FISTA-TV reconstruction algorithm provides higher quality tomograms allowing easier segmentation of the reconstructed volume for a better final processing and analysis.

Optimization of mean-shift scale parameters on the EGEE grid.

This paper studies the optimization of Mean-Shift (MS) image filtering scale parameters. A parameter sweep experiment representing 164 days of CPU is performed on the EGEE grid. The mathematical foundations of Mean-Shift and the grid environment used for the deployment are described in details. The experiments and results are then discussed highlighting the efficiency of gradient ascent algorithm for MS parameters optimization and a number of grid observations related to data transfers, reliability, task scheduling, CPU time and usability.