Search

Visual inspection of microscopic samples is still the gold standard diagnostic methodology for many global health diseases. Soil-transmitted helminth infection affects 1.5 billion people worldwide, and is the most prevalent disease among the Neglected Tropical Diseases. It is diagnosed by manual examination of stool samples by microscopy, which is a time-consuming task and requires trained personnel and high specialization. Artificial intelligence could automate this task making the diagnosis more accessible. Still, it needs a large amount of annotated training data coming from experts.In this work, we proposed the use of crowdsourced annotated medical images to train AI models (neural networks) for the detection of soil-transmitted helminthiasis in microscopy images from stool samples leveraging non-expert knowledge collected through playing a video game. We collected annotations made by both school-age children and adults, and we showed that, although the quality of crowdsourced annotations made by school-age children are sightly inferior than the ones made by adults, AI models trained on these crowdsourced annotations perform similarly (AUC of 0.928 and 0.939 respectively), and reach similar performance to the AI model trained on expert annotations (AUC of 0.932). We also showed the impact of the training sample size and continuous training on the performance of the AI models.In conclusion, the workflow proposed in this work combined collective and artificial intelligence for detecting soil-transmitted helminthiasis. Embedded within a digital health platform can be applied to any other medical image analysis task and contribute to reduce the burden of disease.

This work is framed within the general objective of helping to reduce the cost of telepathology in developing countries and rural areas with no access to automated whole slide imaging (WSI) scanners. We present an automated software pipeline to the problem of mosaicing images acquired with a smartphone, attached to a portable, low-cost, robotic microscopic scanner fabricated using 3D printing technology. To achieve this goal, we propose a robust and automatic workflow, which solves all necessary steps to obtain a stitched image, covering the area of interest, from a set of initial 2D grid of overlapping images, including vignetting correction, lens distortion correction, registration and blending. Optimized solutions, like Voronoi cells and Laplacian blending strategies, are adapted to the low-cost optics and scanner device, and solve imperfections caused using smartphone camera optics. The presented solution can obtain histopathological virtual slides with diagnostic value using a low-cost portable device.

A new workflow is proposed to update the intraoperative electron radiotherapy (IOERT) planning refreshing the position and orientation (pose) of a virtual applicator with respect to the preoperative computed tomography (CT) with the actual pose during surgery. The workflow proposed relies on a robust registration of the preoperative CT and intraoperative projection radiographs acquired with a C-arm system. The workflow initially performs a geometric calibration of the C-arm using fiducials placed on the applicator. In the next step, a point-based 2D–3D registration based on fiducials positioned on the patient’s skin is performed, followed by an intensity-based registration that refines the point-based registration result. The performance of the workflow has been evaluated using a realistic physical phantom consisting of a pig lower limb and its corresponding CT and 7 C-arm projections at different poses. The accuracy has been measured with respect to the applicator origin and axis before and after the registration refinement process. A feasibility study with human data is also included. Error analysis revealed angular accuracy of 0.9 ± 0.7 degrees and translational accuracy of 1.9 ± 1 mm. Our experiments demonstrated that the proposed workflow can achieve subdegree angular accuracy in locating the applicator with respect to the preoperative CT to update and supervise the IOERT planning right before radiation delivery. The proposed workflow could be easily implementable in a routine, corresponding to a significant improvement in quality assurance during IOERT procedures.

A new workflow is proposed to update the intraoperative electron radiotherapy (IOERT) planning refreshing the position and orientation (pose) of a virtual applicator with respect to the preoperative computed tomography (CT) with the actual pose during surgery. The workflow proposed relies on a robust registration of the preoperative CT and intraoperative projection radiographs acquired with a C-arm system. The workflow initially performs a geometric calibration of the C-arm using fiducials placed on the applicator. In the next step, a point-based 2D-3D registration based on fiducials positioned on the patient's skin is performed, followed by an intensity-based registration that refines the point-based registration result. The performance of the workflow has been evaluated using a realistic physical phantom consisting of a pig lower limb and its corresponding CT and 7 C-arm projections at different poses. The accuracy has been measured with respect to the applicator origin and axis before and after the registration refinement process. A feasibility study with human data is also included. Error analysis revealed angular accuracy of 0.9 ± 0.7 degrees and translational accuracy of 1.9 ± 1 mm. Our experiments demonstrated that the proposed workflow can achieve subdegree angular accuracy in locating the applicator with respect to the preoperative CT to update and supervise the IOERT planning right before radiation delivery. The proposed workflow could be easily implementable in a routine, corresponding to a significant improvement in quality assurance during IOERT procedures.

Paraseptal emphysema (PSE) is a relatively unexplored emphysema subtype that is usually asymptomatic, but recently associated with interstitial lung abnormalities which are related with clinical outcomes, including mortality. Previous local-based methods for emphysema subtype quantification do not properly characterize PSE. This is in part for their inability to properly capture the global aspect of the disease, as some the PSE lesions can involved large regions along the chest wall. It is our assumption, that path-based approaches are not well-suited to identify this subtype and segmentation is a better paradigm. In this work we propose and introduce the Slice-Recovery network (SR-Net) that leverages 3D contextual information for 2D segmentation of PSE lesions in CT images. For that purpose, a novel convolutional network architecture is presented, which follows an encoding-decoding path that processes a 3D volume to generate a 2D segmentation map. The dataset used for training and testing the method comprised 664 images, coming from 111 CT scans. The results demonstrate the benefit of the proposed approach which incorporate 3D context information to the network and the ability of the proposed method to identify and segment PSE lesions with different sizes even in the presence of other emphysema subtypes in an advanced stage.

In this article we propose and validate a fully automatic tool for emphysema classification in Computed Tomography (CT) images. We hypothesize that a relatively simple Convolutional Neural Network (CNN) architecture can learn even better discriminative features from the input data compared with more complex and deeper architectures. The proposed architecture is comprised of only 4 convolutional and 3 pooling layers, where the input corresponds to a 2.5D multiview representation of the pulmonary segment tissue to classify, corresponding to axial, sagittal and coronal views. The proposed architecture is compared to similar 2D CNN and 3D CNN, and to more complex architectures which involve a larger number of parameters (up to six times larger). This method has been evaluated in 1553 tissue samples, and achieves an overall sensitivity of 81.78 % and a specificity of 97.34%, and results show that the proposed method outperforms deeper state-of-the-art architectures particularly designed for lung pattern classification. The method shows satisfactory results in full-lung classification.

Density masking is the de-facto quantitative imaging phenotype for emphysema that is widely used by the clinical community. Density masking defines the burden of emphysema by a fixed threshold, usually between −910 HU and −950 HU, that has been experimentally validated with histology. In this work, we formalized emphysema quantification by means of statistical inference. We show that a non-central Gamma is a good approximation for the local distribution of image intensities for normal and emphysema tissue. We then propose a test statistic in terms of the sample mean of a truncated non-central Gamma random variable. Our results show that this approach is well-suited for the detection of emphysema and superior to standard density masking. The statistical method was tested in a dataset of 1337 samples obtained from 9 different scanner models in subjects with COPD. Results showed an increase of 17% when compared to the density masking approach, and an overall accuracy of 94.09%.

Optical Coherence Tomography (OCT) has shown a great potential as a complementary imaging tool in the diagnosis of skin diseases. Speckle noise is the most prominent artifact present in OCT images and could limit the interpretation and detection capabilities. In this work we evaluate various denoising filters with high edge-preserving potential for the reduction of speckle noise in 256 dermatological OCT B-scans. Our results show that the Enhanced Sigma Filter and the Block Matching 3-D (BM3D) as 2D denoising filters and the Wavelet Multiframe algorithm considering adjacent B-scans achieved the best results in terms of the enhancement quality metrics used. Our results suggest that a combination of 2D filtering followed by a wavelet based compounding algorithm may significantly reduce speckle, increasing signal-to-noise and contrast-to-noise ratios, without the need of extra acquisitions of the same frame.

Automated medical image analysis requires methods to localize anatomic structures in the presence of normal interpatient variability, pathology, and the different protocols used to acquire images for different clinical settings. Recent advances have improved object detection in the context of natural images, but they have not been adapted to the 3D context of medical images. In this paper we present a 2.5D object detector designed to locate, without any user interaction, the left and right heart ventricles in Computed Tomography Pulmonary Angiography (CTPA) images. A 2D object detector is trained to find ventricles on axial slices. Those detections are automatically clustered according to their size and position. The cluster with highest score, representing the 3D location of the ventricle, is then selected. The proposed method is validated in 403 CTPA studies obtained in patients with clinically suspected pulmonary embolism. Both ventricles are properly detected in 94.7% of the cases. The proposed method is very generic and can be easily adapted to detect other structures in medical images.

The understanding of the embryogenesis in living systems requires reliable quantitative analysis of the cell migration throughout all the stages of development. This is a major challenge of the ”in-toto” reconstruction based on different modalities of ”in-vivo” imaging techniques-spatio-temporal resolution and image artifacts and noise. Several methods for cell tracking are available, but expensive manual interaction-time and human resources — is always required to enforce coherence. Because of this limitation it is necessary to restrict the experiments or assume an uncontrolled error rate. Is it possible to obtain automated reliable measurements of migration? can we provide a seed for biologists to complete cell lineages efficiently? We propose a filtering technique that considers trajectories as spatio-temporal connected structures that prunes out those that might introduce noise and false positives by using multi-dimensional morphological operators.

This work presents the application of an updated version of an existing Monte Carlo (MC) code for photon and electron transport to intraoperative radiotherapy (IORT) dose planning. The code has been optimized to provide an estimation of the dose delivered by the linear accelerator (LINAC) in very short time and has been incorporated as a submodule of an advanced IORT planning system under development, which will enable far more accurate pre-planning as well as the intraoperative treatment modifications considering actual information taken at the operating room (OR) during surgery, for instance with a mobile C-arm.

A description of mitral valve function requires both an accurate anatomical description of the three-dimensional (3D) structure as well as a description of the leaflets' rapid motion. Current two-dimensional (2D) ultrasound imaging is unable to capture the 3D anatomy and 3D imaging suffers from severely reduced temporal resolution. We present a method to utilize two sequential 3D transesophageal echocardiographic (TEE) ultrasound acquisitions with a temporal delay to create a dataset that captures the 3D anatomical description of the valve and has a higher temporal resolution. The enhanced dataset should provide better information for modeling the motion of the mitral vale.

An experimental structured light projection system which includes a miniaturized projector is described. The system has been designed to be integrated in a fluorescence molecular tomography (FMT) prototype in order to reconstruct the surface of mice and phantom studies. A high-accuracy phase map is retrieved with phase-shifted sinusoidal fringes. Phase error due to the nonlinear gamma function of the pico-projector is calibrated and compensated. Robust phase unwrapping is performed with an additional Gray-code projection sequence. An automatic phase-to-height non-linear calibration scheme has been applied using objects located in the extremes of the field of view. The accuracy of the proposed method has been tested with a realistic mouse model and ray-tracing software.

In this work we propose an active surface method to segment complete liver volumes from preoperative CT abdominal images. The method finds the surface that minimizes an energy function combining intensity inside and outside the surface, gradient information and curvature restrictions. The implementation is based on a level set technique following a multi-resolution strategy to reduce computing time. It requires only a single seed point inside the liver to initialize the active surface. The algorithm has been validated on a set of previously diagnosed livers. Resulting segmentations have been supervised by clinicians and radiologists, and numerically evaluated in terms of volume measurements with respect to those obtained from radiologists' manual segmentations. Additionally, radiologists analyzed the necessity of additional corrections on segmenting volumes.

Proceeding of: 2009 IEEE Nuclear Science Symposium Conference Record (NSS/MIC), Orlando, Florida, 25-31 October 2009 Monte Carlo methods provide a flexible and rigorous solution to the problem of light transport in turbid media, which enable approaching complex geometries for a closed analytical solution is not feasible. The simulator implements local rules of propagation in the form of probability density functions that depend on the local optical properties of the tissue. This work presents a flexible simulator that can be applied in multiple applications related to optical tomography. In particular, unlike previous codes, the simulator explicitly supports fluorescent-tissues and variance reduction and code parallelization techniques are implemented in order to speed up the execution with fluorochrome-labelled agents. The simulator is validated with simple geometries for which an analytical solution exists, as well as with an experimental polyester resin based optical phantom. This work was supported in part by the Ministry of Science and Innovation under projects TEC2008-06715 and TEC2007-64731/TCM and by the EU’s 7th Frame Programme under contract HEALTH-F5-2008-201792

In this work we propose an image registration algorithm to automatically fit protein atomic domain models into medium-resolution three-dimensional electron microscopy reconstructions (3D-EM map). The approach employs a flexible registration algorithm whose optimizer controls the generation of stereo-chemically correct models from a given reference domain belonging to a super-family of proteins. The proposed algorithm generates models automatically in the correct direction until the optimizer converges to the best fitted model. A local gradient optimizer is employed for this task. Mutual Information is used as an alternative to Cross Correlation Coefficient. An additional rigid registration that uses a local gradient optimization is applied between each generated model and the 3D-EM map.

Positron emission tomography (PET) requires accurate timing of scintillation events to properly discriminate between coincident and noncoincident pairs. The traditional solution to timing is based on custom application specific integrated circuits (ASIC) designs, whose cost may not be justified in the design of experimental small animal PET scanners. The new generation of PET scanners introduces the idea of continuous sampling of the detected scintillation pulse, replacing event-triggered acquisition front-ends. This approach enables new options to the timing procedure based on digital processing of the sampled pulse signal. This work proposes a time stamping algorithm based on the optically matched filter and compares the potential performance benefits of this approach versus other FIR-based timing algorithms, some of which have been already implemented by different authors. Results show that the coincidence timing resolution may be as low as 1.5 ns without the need of expensive high-speed converters when the proper signal processing is applied IEEE Nuclear and Plasma Sciences Society Publicado

The objective of this work is to compare the results obtained from the motion analysis of tagged vs. CINE MR sequences when using spatio-temporal non-rigid registration techniques based on pixel intensity. Those techniques have been previously validated on tagged MR images. Moreover, registration algorithms have been applied to MR CINE sequences to obtain radial displacement and strain parameters demonstrating its usefulness to quantify regional myocardial function. Tagged and CINE MR short axis sequences from 10 subjects were examined. Four segments were manually selected in both the tagged MR and CINE sequences. Automatic estimation of the myocardial motion field was performed using a consecutive non-rigid registration algorithm based on a semilocal Bspline parametric model. Finally, a statistical analysis was applied to compare the systolic displacement and strain estimations from both types of sequences. An important discrepancy between results obtained from tagged MR based strain analysis and CINE MR has been found.

Proceeding of: 2007 IEEE Nuclear Science Symposium Conference Record (NSS'07), Honolulu, Hawaii, USA, Oct. 27 - Nov. 3, 2007 Small animal PET systems based on a reduced number of rotating planar detectors have advantages in terms of cost/performance for high resolution imaging. The huge integration capability of current semiconductor technologies has favoured miniaturization and popularised the concept of system on a chip. Following this trend, this work attempts to integrate in a single chip the acquisition electronics and coincidence unit of a fully digital four-head PET scanner. The proof of concept is based on commercially available industrial modules, including a FPGAbased high speed acquisition board combined with a rackable computer, both in 6U cPCI form. The four acquisition digital blocks, coincidence logic and communication interface take up close to 30% of the available resources of the FPGA, providing margin for improvement in the system under development. This work has been partially funded by the Spanish Ministry of Education and Science through the FPU grant program as well as by the CDTEAM consortium as part of the CENIT Spanish research plan.

Respiratory motion in emission tomography leads to reduced image quality. Proposed correction methodology has been concentrating on the use of respiratory synchronised acquisitions leading to gated frames. Such frames however are of low signal to noise ratio as a result of containing reduced statistics. Therefore a method accounting for respiratory motion effects without affecting the statistical quality of the reconstructed images is necessary. In this work we describe the implementation of an elastic transformation within a list-mode based reconstruction for the correction of respiratory motion over the thorax. The developed algorithm was evaluated using datasets of the NCAT phantom generated at different points throughout the respiratory cycle. List mode data based PET simulated frames were subsequently produced by combining the NCAT datasets with a Monte Carlo simulation. Transformation parameters accounting for respiratory motion were estimated according to an elastic registration of the NCAT dynamic CT images and were subsequently applied during the image reconstruction of the original emission list mode data. The One-pass list mode EM (OPL-EM) algorithm was modified to integrate the elastic transformation. The corrected images were compared with those produced using an affine transformation of list mode data prior to reconstruction, as well as with uncorrected respiratory motion average images. Results demonstrate that although both correction techniques considered lead to significant improvements in accounting for respiratory motion artefacts in the lungs and heart, the elastic transformation based correction leads to a more uniform improvement across the lung field for different lesion sizes.

Respiratory motion in emission tomography leads to reduced image quality. Proposed correction methodology has been concentrating on the use of respiratory synchronised acquisitions leading to gated frames. Such frames however are of low signal to noise ratio as a result of containing reduced statistics. Therefore a method accounting for respiratory motion effects without affecting the statistical quality of the reconstructed images is necessary. In this work we describe the implementation of an elastic transformation within a list-mode based reconstruction for the correction of respiratory motion over the thorax. The developed algorithm was evaluated using datasets of the NCAT phantom generated at different points throughout the respiratory cycle. List mode data based PET simulated frames were subsequently produced by combining the NCAT datasets with a Monte Carlo simulation. Transformation parameters accounting for respiratory motion were estimated according to an elastic registration of the NCAT dynamic CT images and were subsequently applied during the image reconstruction of the original emission list mode data. The One-pass list mode EM (OPL-EM) algorithm was modified to integrate the elastic transformation in the sensitivity matrix. Three different implementations have been investigated (no interpolation, trilinear interpolation, b-spline functions incorporation). The corrected images were compared with uncorrected respiratory motion average images. Results demonstrate that the use of elastic transformations in the reconstruction system matrix lead to uniform improvement across the lung field for different lesion sizes. The use of a trilinear interpolation or the incorporation of the b-spline functions lead to times of execution equivalent to standard image reconstruction. However, trilinear interpolation leads to artefacts in areas such as the diaphragm where the largest elastic deformations are occurring.