Search

Abstract Reliable automatic diagnosis of Diabetic Retinopathy (DR) and Macular Edema (ME) is an invaluable asset in improving the rate of monitored patients among at-risk populations and in enabling earlier treatments before the pathology progresses and threatens vision. However, the explainability of screening models is still an open question, and specifically designed datasets are required to support the research. We present MAPLES-DR (MESSIDOR Anatomical and Pathological Labels for Explainable Screening of Diabetic Retinopathy), which contains, for 198 images of the MESSIDOR public fundus dataset, new diagnoses for DR and ME as well as new pixel-wise segmentation maps for 10 anatomical and pathological biomarkers related to DR. This paper documents the design choices and the annotation procedure that produced MAPLES-DR, discusses the interobserver variability and the overall quality of the annotations, and provides guidelines on using the dataset in a machine learning context.

Abstract The identification of eye diseases and their progression often relies on a clear visualization of the anatomy and on different metrics extracted from Optical Coherence Tomography (OCT) B-scans. However, speckle noise hinders the quality of rapid OCT imaging, hampering the extraction and reliability of biomarkers that require time series. By synchronizing the acquisition of OCT images with the timing of the cardiac pulse, we transform a low-quality OCT video into a clear version by phase-wrapping each frame to the heart pulsation and averaging frames that correspond to the same instant in the cardiac cycle. Here, we compare the performance of our one-cycle denoising strategy with a deep-learning architecture, Noise2Noise, as well as classical denoising methods such as BM3D and Non-Local Means (NLM). We systematically analyze different image quality descriptors as well as region-specific metrics to assess the denoising performance based on the anatomy of the eye. The one-cycle method achieves the highest denoising performance, increases image quality and preserves the high-resolution structures within the eye tissues. The proposed workflow can be readily implemented in a clinical setting.

Automatic and accurate dental arch segmentation is a fundamental task in computer-aided dentistry. Recent trends in digital dentistry are tackling the design of 3D crowns using artificial intelligence, which initially requires a proper semantic segmentation of teeth from intraoral scans (IOS). In practice, most IOS are partial with as few as three teeth on the scanned arch, and some of them might have preparations, missing, or incomplete teeth. Existing deep learning-based methods (e.g., MeshSegNet, DArch) were proposed for dental arch segmentation, but they are not as efficient for partial arches that include imperfections such as missing teeth and preparations. In this work, we present the ArchSeg framework that can leverage various deep learning models for semantic segmentation of perfect and imperfect dental arches. The Point Transformer V2 deep learning model is used as the backbone for the ArchSeg framework. We present experiments to demonstrate the efficiency of the proposed framework to segment arches with various types of imperfections. Using a raw dental arch scan with two labels indicating the range of present teeth in the arch (i.e., the first and the last teeth), our ArchSeg can segment a standalone dental arch or a pair of aligned master/antagonist arches with more available information (i.e., die mesh). Two generic models are trained for lower and upper arches; they achieve dice similarity coefficient scores of 0.936±0.008 and 0.948±0.007, respectively, on test sets composed of challenging imperfect arches. Our work also highlights the impact of appropriate data pre-processing and post-processing on the final segmentation performance. Our ablation study shows that the segmentation performance of the Point Transformer V2 model integrated in our framework is improved compared with the original standalone model.

Uncertainty estimations through approximate Bayesian inference provide interesting insights to deep neural networks' behavior. In unsupervised learning tasks, where expert labels are unavailable, it becomes ever more important to critique the model through uncertainties. This paper presents a proof-of-concept for generalizing the aleatoric and epistemic uncertainties in unsupervised MR-CT synthesis of scoliotic spines. A novel adaptation of the cycle-consistency constraint in CycleGAN is proposed such that the model predicts the aleatoric uncertainty maps in addition to the standard volume-to-volume translation between Magnetic Resonance (MR) and Computed Tomography (CT) data. Ablation experiments were performed to understand uncertainty estimation as an implicit regularizer and a measure of the model's confidence. The aleatoric uncertainty helps in distinguishing between the bone and soft-tissue regions in CT and MR data during translation, while the epistemic uncertainty provides interpretable information to the user for downstream tasks.

During a crown generation procedure, dental technicians depend on commercial software to generate a margin line to define the design boundary for the crown. The margin line generation remains a non-reproducible, inconsistent, and challenging procedure. In this work, we propose to generate margin line points on prepared teeth meshes using adaptive point learning inspired by the AdaPointTr model. We extracted ground truth margin lines as point clouds from the prepared teeth and crown bottom meshes. The chamfer distance (CD) and infoCD loss functions were used for training a supervised deep learning model that outputs a margin line as a point cloud. To enhance the generation results, the deep learning model was trained based on three different resolutions of the target margin lines, which were used to back-propagate the losses. Five folds were trained and an ensemble model was constructed. The training and test sets contained 913 and 134 samples, respectively, covering all teeth positions. Intraoral scanning was used to collect all samples. Our post-processing involves removing outlier points based on local point density and principal component analysis (PCA) followed by a spline prediction. Comparing our final spline predictions with the ground truth margin line using CD, we achieved a median distance of 0.137 mm. The median Hausdorff distance was 0.242 mm. We also propose a novel confidence metric for uncertainty quantification of generated margin lines during deployment. The metric was defined based on the percentage of removed outliers during the post-processing stage. The proposed end-to-end framework helps dental professionals in generating and evaluating margin lines consistently. The findings underscore the potential of deep learning to revolutionize the detection and extraction of 3D landmarks, offering personalized and robust methods to meet the increasing demands for precision and efficiency in the medical field.

Objective: To develop a noninvasive technique to quantitatively assess the pulsatile deformation due to cardiac contractions of the optic nerve head (ONH). Design: Evaluation of a diagnostic test or technology. Participants: Healthy subjects with no history of refractive surgery, divided into 2 cohorts on the basis of their axial length (AL). Methods: We present a noninvasive technique to quantitatively assess the pulsatile deformation of the ONH tissue by combining high-frequency OCT imaging and widely available image processing algorithms. We performed a thorough validation of the approach, numerically and experimentally, evaluating the sensitivity of the method to artificially induced deformation and its robustness to different noise levels. We performed deformation measurements in cohorts of healthy (n = 9) and myopic (n = 5) subjects in different physiological strain conditions by calculating the amplitude of tissue displacement in both the primary position and abduction. The head rotation was measured using a goniometer. During imaging in abduction, the head was rotated 40° ± 3°, and subjects were instructed to direct their gaze toward the OCT visual target. Main Outcome Measures: Pulsatile tissue displacement maps. Results: The robustness of the method was assessed using artificial deformations and increasing noise levels. The results show acceptable absolute errors before the noise simulations grossly exaggerate image degradation. For the group of subjects with AL of < 25 mm (n = 9), the median pulsatile displacement of the ONH was 7.8 ± 1.3 μm in the primary position and 8.9 ± 1.2 μm in abduction. The Wilcoxon test showed a significant difference (P ≤ 0.005) between the 2 paired measures. Reproducibility was tested in 2 different sessions in 5 different subjects with the same intraocular pressure, and an intraclass correlation coefficient of 0.99 was obtained (P < 0.005). Conclusions: The computational pipeline demonstrated good reproducibility and had the capacity to accurately map the pulsatile deformation of the optic nerve. In a clinical setting, we detected physiological changes in normal subjects supporting its translation potential as a novel biomarker for the diagnosis and progression of optic nerve diseases.

Abstract Recent studies suggested that cerebrovascular micro-occlusions, i.e. microstokes, could lead to ischemic tissue infarctions and cognitive deficits. Due to their small size, identifying measurable biomarkers of these microvascular lesions remains a major challenge. This work aims to simulate potential MRI signatures combining arterial spin labeling (ASL) and multi-directional diffusion-weighted imaging (DWI). Driving our hypothesis are recent observations demonstrating a radial reorientation of microvasculature around the micro-infarction locus during recovery in mice. Synthetic capillary beds, randomly- and radially-oriented, and optical coherence tomography (OCT) angiograms, acquired in the barrel cortex of mice (n = 5) before and after inducing targeted photothrombosis, were analyzed. Computational vascular graphs combined with a 3D Monte-Carlo simulator were used to characterize the magnetic resonance (MR) response, encompassing the effects of magnetic field perturbations caused by deoxyhemoglobin, and the advection and diffusion of the nuclear spins. We quantified the minimal intravoxel signal loss ratio when applying multiple gradient directions, at varying sequence parameters with and without ASL. With ASL, our results demonstrate a significant difference (p < 0.05) between the signal-ratios computed at baseline and 3 weeks after photothrombosis. The statistical power further increased (p < 0.005) using angiograms measured at week 4. Without ASL, no reliable signal change was found. We found that higher ratios, and accordingly improved significance, were achieved at lower magnetic field strengths (e.g., B0 = 3T) and shorter echo time TE (< 16 ms). Our simulations suggest that microstrokes might be characterized through ASL-DWI sequence, providing necessary insights for posterior experimental validations, and ultimately, future translational trials.

Acute respiratory distress syndrome (ARDS), including severe pulmonary COVID infection, is associated with a high mortality rate. It is crucial to detect ARDS early, as a late diagnosis may lead to serious complications in treatment. One of the challenges in ARDS diagnosis is chest X-ray (CXR) interpretation. ARDS causes diffuse infiltrates through the lungs that must be identified using chest radiography. In this paper, we present a web-based platform leveraging artificial intelligence (AI) to automatically assess pediatric ARDS (PARDS) using CXR images. Our system computes a severity score to identify and grade ARDS in CXR images. Moreover, the platform provides an image highlighting the lung fields, which can be utilized for prospective AI-based systems. A deep learning (DL) approach is employed to analyze the input data. A novel DL model, named Dense-Ynet, is trained using a CXR dataset in which clinical specialists previously labelled the two halves (upper and lower) of each lung. The assessment results show that our platform achieves a recall rate of 95.25% and a precision of 88.02%. The web platform, named PARDS-CxR, assigns severity scores to input CXR images that are compatible with current definitions of ARDS and PARDS. Once it has undergone external validation, PARDS-CxR will serve as an essential component in a clinical AI framework for diagnosing ARDS.

Minimally invasive surgeries (MISs) are gaining popularity as alternatives to conventional open surgeries. In thoracoscopic scoliosis MIS, fluoroscopy is used to guide pedicle screw placement and to visualise the effect of the intervention on the spine curvature. However, cosmetic external appearance is the most important concern for patients, while correction of the spine and achieving coronal and sagittal trunk balance are the top priorities for surgeons. The authors present the feasibility study of the first intra-operative assistive system for scoliosis surgery composed of a single RGBD camera affixed on a C-arm which allows visualising in real time the surgery effects on the patient trunk surface in the transverse plane. They perform three feasibility experiments from simulated data based on scoliotic patients to live acquisition from non-scoliotic mannequin and person, all showing that the proposed system accuracy is comparable with scoliotic surface reconstruction state of art.

Age-related macular degeneration (AMD) is a disease which causes visual deficiency and irreversible blindness to the elderly. In this paper, an automatic classification method for AMD is proposed to perform robust and reproducible assessments in a telemedicine context. First, a study was carried out to highlight the most relevant features for AMD characterization based on texture, color, and visual context in fundus images. A support vector machine and a random forest were used to classify images according to the different AMD stages following the AREDS protocol and to evaluate the features’ relevance. Experiments were conducted on a database of 279 fundus images coming from a telemedicine platform. The results demonstrate that local binary patterns in multiresolution are the most relevant for AMD classification, regardless of the classifier used. Depending on the classification task, our method achieves promising performances with areas under the ROC curve between 0.739 and 0.874 for screening and between 0.469 and 0.685 for grading. Moreover, the proposed automatic AMD classification system is robust with respect to image quality.

Mechanical ventilation is a very effective therapy, but with many complications. Simulators are used in many fields, including medicine, to enhance safety issues. In the intensive care unit, they are used for teaching cardiorespiratory physiology and ventilation, for testing ventilator performance, for forecasting the effect of ventilatory support, and to determine optimal ventilatory management. They are also used in research and development of clinical decision support systems (CDSSs) and explicit computerized protocols in closed loop. For all those reasons, cardiorespiratory simulators are one of the tools that help to decrease mechanical ventilation duration and complications. This paper describes the different types of simulators described in the literature for physiologic simulation and modeling of the respiratory system, including a new simulator (SimulResp), and proposes a validation process for these simulators.

Articulated models are commonly used for recognition tasks in robotics and in gait analysis, but can also be extremely useful to develop analytical methods targeting spinal deformities studies. The threedimensional analysis of these deformities is critical since they are complex and not restricted to a given plane. Thus, they cannot be assessed as a two-dimensional phenomenon. However, analyzing large databases of 3D spine models is a difficult and time-consuming task. In this context, a method that automatically extracts the most important deformation modes from sets of articulated spine models is proposed. The spine was modeled with two levels of details. In the first level, the global shape of the spine was expressed using a set of rigid transformations that superpose local coordinates systems of neighboring vertebrae. In the second level, anatomical landmarks measured with respect to a vertebra’s local coordinate system were used to quantify vertebra shape. These articulated spine models do not naturally belong to a vector space because of the vertebral rotations. The Fréchet mean, which is a generalization of the conventional mean to Riemannian manifolds, was thus used to compute the mean spine shape. Moreover, a generalized covariance computed in the tangent space of the Fréchet mean was used to construct a statistical shape model of the scoliotic spine. The principal deformation modes were then extracted by performing a principal component analysis (PCA) on the generalized covariance matrix. The principal deformations modes were computed for a large database of untreated scoliotic patients. The obtained results indicate that combining rotation, translation and local vertebra shape into a unified framework leads to an effective and meaningful analysis method for articulated anatomical structures. The computed deformation modes also revealed clinically relevant information. For instance, the first mode of deformation is associated with patients’ growth, the second is a double thoraco-lumbar curve and the third is a thoracic curve. Other experiments were performed on patients classified by orthopedists with respect to a widely used two-dimensional surgical planning system (the Lenke classification) and patterns relevant to the definition of a new three-dimensional classification were identified. Finally, relationships between local vertebrae shapes and global spine shape (such as vertebra wedging) were demonstrated using a sample of 3D spine reconstructions with 14 anatomical landmarks per vertebra. KeyWords: Shape Analysis, Articulated Models, Spinal Deformities, Scoliosis, 3D Reconstruction, Surgical Classifications.

Abstract Background For some scoliotic patients the spinal instrumentation is inevitable. Among these patients, those with stiff curvature will need thoracoscopic disk resection. The removal of the intervertebral disk with only thoracoscopic images is a tedious and challenging task for the surgeon. With computer aided surgery and 3D visualisation of the interverterbral disk during surgery, surgeons will have access to additional information such as the remaining disk tissue or the distance of surgical tools from critical anatomical structures like the aorta or spinal canal. We hypothesized that automatically extracting 3D information of the intervertebral disk from MR images would aid the surgeons to evaluate the remaining disk and would add a security factor to the patient during thoracoscopic disk resection. Methods This paper presents a quantitative evaluation of an automatic segmentation method for 3D reconstruction of intervertebral scoliotic disks from MR images. The automatic segmentation method is based on the watershed technique and morphological operators. The 3D Dice Similarity Coefficient (DSC) is the main statistical metric used to validate the automatically detected preoperative disk volumes. The automatic detections of intervertebral disks of real clinical MR images are compared to manual segmentation done by clinicians. Results Results show that depending on the type of MR acquisition sequence, the 3D DSC can be as high as 0.79 (±0.04). These 3D results are also supported by a 2D quantitative evaluation as well as by robustness and variability evaluations. The mean discrepancy (in 2D) between the manual and automatic segmentations for regions around the spinal canal is of 1.8 (±0.8) mm. The robustness study shows that among the five factors evaluated, only the type of MRI acquisition sequence can affect the segmentation results. Finally, the variability of the automatic segmentation method is lower than the variability associated with manual segmentation performed by different physicians. Conclusions This comprehensive evaluation of the automatic segmentation and 3D reconstruction of intervertebral disks shows that the proposed technique used with specific MRI acquisition protocol can detect intervertebral disk of scoliotic patient. The newly developed technique is promising for clinical context and can eventually help surgeons during thoracoscopic intervertebral disk resection.