Natasha Lepore

Improvement of Bragg peak shift estimation using dimensionality reduction techniques and predictive linear modeling.

With the emergence of clinical prototypes and first patient acquisitions for proton therapy, the research on prompt gamma imaging is aiming at making most use of the prompt gamma data for in vivo estimation of any shift from expected Bragg peak (BP). The simple problem of matching the measured prompt gamma profile of each pencil beam with a reference simulation from the treatment plan is actually made complex by uncertainties which can translate into distortions during treatment. We will illustrate this challenge and demonstrate the robustness of a predictive linear model we proposed for BP shift estimation based on principal component analysis (PCA) method. It considered the first clinical knife-edge slit camera design in use with anthropomorphic phantom CT data. Particularly, 4115 error scenarios were simulated for the learning model. PCA was applied to the training input randomly chosen from 500 scenarios for eliminating data collinearities. A total variance of 99.95% was used for representing the testing input from 3615 scenarios. This model improved the BP shift estimation by an average of 63±19% in a range between -2.5% and 86%, comparing to our previous profile shift (PS) method. The robustness of our method was demonstrated by a comparative study conducted by applying 1000 times Poisson noise to each profile. 67% cases obtained by the learning model had lower prediction errors than those obtained by PS method. The estimation accuracy ranged between 0.31 ± 0.22 mm and 1.84 ± 8.98 mm for the learning model, while for PS method it ranged between 0.3 ± 0.25 mm and 20.71 ± 8.38 mm.

Acute effect of Vagus nerve stimulation parameters on cardiac chronotropic, inotropic, and dromotropic responses.

Vagus nerve stimulation (VNS) is an established therapy for drug-resistant epilepsy and depression, and is considered as a potential therapy for other pathologies, including Heart Failure (HF) or inflammatory diseases. In the case of HF, several experimental studies on animals have shown an improvement in the cardiac function and a reverse remodeling of the cardiac cavity when VNS is applied. However, recent clinical trials have not been able to reproduce the same response in humans. One of the hypothesis to explain this lack of response is related to the way in which stimulation parameters are defined. The combined effect of VNS parameters is still poorly-known, especially in the case of VNS synchronously delivered with cardiac activity. In this paper, we propose a methodology to analyze the acute cardiovascular effects of VNS parameters individually, as well as their interactive effects. A Latin hypercube sampling method was applied to design a uniform experimental plan. Data gathered from this experimental plan was used to produce a Gaussian process regression (GPR) model in order to estimate unobserved VNS sequences. Finally, a Morris screening sensitivity analysis method was applied to each obtained GPR model. Results highlight dominant effects of pulse current, pulse width and number of pulses over frequency and delay and, more importantly, the degree of interactions between these parameters on the most important acute cardiovascular responses. In particular, high interacting effects between current and pulse width were found. Similar sensitivity profiles were observed for chronotropic, dromotropic and inotropic effects. These findings are of primary importance for the future development of closed-loop, personalized neuromodulator technologies.

Detection of the default mode network by an anisotropic analysis.

This document presents a proposal devoted to improve the detection of the default mode network (DMN) in resting state functional magnetic resonance imaging in noisy conditions caused by head movement. The proposed approach is inspired by the hierarchical treatment of information, in particular at the level of the brain basal ganglia. Essentially, the fact that information must be selected and reduced suggests propagation of information in the Central Nervous System (CNS) is anisotropic. Under this hypothesis, the reconstruction of information of activation should follow an anisotropic pattern. In this work, an anisotropic filter is used to recover the DMN that is perturbed by simulated motion artifacts. Results obtained show this approach outperforms the state-of-the-art methods by 5.93% PSNR.

Semantic knowledge for histopathological image analysis: from ontologies to processing portals and deep learning.

This article presents our vision about the next generation of challenges in computational/digital pathology. The key role of the domain ontology, developed in a sustainable manner (i.e. using reference checklists and protocols, as the living semantic repositories), opens the way to effective/sustainable traceability and relevance feedback concerning the use of existing machine learning algorithms, proven to be very performant in the latest digital pathology challenges (i.e. convolutional neural networks). Being able to work in an accessible web-service environment, with strictly controlled issues regarding intellectual property (image and data processing/analysis algorithms) and medical data/image confidentiality is essential for the future. Among the web-services involved in the proposed approach, the living yellow pages in the area of computational pathology seems to be very important in order to reach an operational awareness, validation, and feasibility. This represents a very promising way to go to the next generation of tools, able to bring more guidance to the computer scientists and confidence to the pathologists, towards an effective/efficient daily use. Besides, a consistent feedback and insights will be more likely to emerge in the near future - from these sophisticated machine learning tools - back to the pathologists--, strengthening, therefore, the interaction between the different actors of a sustainable biomedical ecosystem (patients, clinicians, biologists, engineers, scientists etc.). Beside going digital/computational - with virtual slide technology demanding new workflows--, Pathology must prepare for another coming revolution: semantic web technologies now enable the knowledge of experts to be stored in databases, shared through the Internet, and accessible by machines. Traceability, disambiguation of reports, quality monitoring, interoperability between health centers are some of the associated benefits that pathologists were seeking. However, major changes are also to be expected for the relation of human diagnosis to machine based procedures. Improving on a former imaging platform which used a local knowledge base and a reasoning engine to combine image processing modules into higher level tasks, we propose a framework where different actors of the histopathology imaging world can cooperate using web services - exchanging knowledge as well as imaging services - and where the results of such collaborations on diagnostic related tasks can be evaluated in international challenges such as those recently organized for mitosis detection, nuclear atypia, or tissue architecture in the context of cancer grading. This framework is likely to offer an effective context-guidance and traceability to Deep Learning approaches, with an interesting promising perspective given by the multi-task learning (MTL) paradigm, distinguished by its applicability to several different learning algorithms, its non- reliance on specialized architectures and the promising results demonstrated, in particular towards the problem of weak supervision--, an issue found when direct links from pathology terms in reports to corresponding regions within images are missing.

Supporting the potential of quantitative ultrasonic techniques for the evaluation of platelet concentration.

This article describes the results obtained by making use of a non-destructive, non-invasive ultrasonic system for the acoustic characterization of bovine plasma rich in platelets using digital signal processing techniques. This study includes computational methods based on acoustic spectrometry estimation and experimental measurements of the speed of sound in blood plasma from different samples analyzed, using an ultrasonic field with resonance frequency of 5 MHz. The results showed that the measurements on ultrasonic signals can contribute to the hematological predictions based on the linear regression model applied to the relationship between experimental ultrasonic parameters calculated and platelet concentration, indicating a growth rate of 1 m/s for each 0.90 x103 platelet per mm3. On the other hand, the attenuation coefficient presented changes of 20% in the platelet concentration using a resolution of 0.057 dB/cm MHz.

Analysis of sEMG signals using discrete wavelet transform for muscle fatigue detection.

The purpose of the present article is to characterize sEMG signals to determine muscular fatigue levels. To do this, the signal is decomposed using the discrete wavelet transform, which offers noise filtering features, simplicity and efficiency. sEMG signals on the forearm were acquired and analyzed during the execution of cyclic muscular contractions in the presence and absence of fatigue. When the muscle fatigues, the sEMG signal shows a more erratic behavior of the signal as more energy is required to maintain the effort levels.

Statistical shape (ASM) and appearance (AAM) models for the segmentation of the cerebellum in fetal ultrasound.

The cerebellum is an important structure to determine the gestational age of the fetus, moreover most of the abnormalities it presents are related to growth disorders. In this work, we present the results of the segmentation of the fetal cerebellum applying statistical shape and appearance models. Both models were tested on ultrasound images of the fetal brain taken from 23 pregnant women, between 18 and 24 gestational weeks. The accuracy results obtained on 11 ultrasound images show a mean Hausdorff distance of 6.08 mm between the manual segmentation and the segmentation using active shape model, and a mean Hausdorff distance of 7.54 mm between the manual segmentation and the segmentation using active appearance model. The reported results demonstrate that the active shape model is more robust in the segmentation of the fetal cerebellum in ultrasound images.

Characterizing brain patterns in conversion from mild cognitive impairment (MCI) to Alzheimer's disease.

Structural Magnetic Resonance (MR) brain images should provide quantitative information about the stage and progression of Alzheimer’s disease. However, the use of MRI is limited and practically reduced to corroborate a diagnosis already performed with neuropsychological tools. This paper presents an automated strategy for extraction of relevant anatomic patterns related with the conversion from mild cognitive impairment (MCI) to Alzheimer’s disease (AD) using T1-weighted MR images. The process starts by representing each of the possible classes with models generated from a linear combination of volumes. The difference between models allows us to establish which are the regions where relevant patterns might be located. The approach searches patterns in a space of brain sulci, herein approximated by the most representative gradients found in regions of interest defined by the difference between the linear models. This hypothesis is assessed by training a conventional SVM model with the found relevant patterns under a leave-one-out scheme. The resultant AUC was 0.86 for the group of women and 0.61 for the group of men.

Retinal blood vessel segmentation in high resolution fundus photographs using automated feature parameter estimation.

Several ophthalmological and systemic diseases are manifested through pathological changes in the properties and the distribution of the retinal blood vessels. The characterization of such alterations requires the segmentation of the vasculature, which is a tedious and time-consuming task that is infeasible to be performed manually. Numerous attempts have been made to propose automated methods for segmenting the retinal vasculature from fundus photographs, although their application in real clinical scenarios is usually limited by their ability to deal with images taken at different resolutions. This is likely due to the large number of parameters that have to be properly calibrated according to each image scale. In this paper we propose to apply a novel strategy for automated feature parameter estimation, combined with a vessel segmentation method based on fully connected conditional random fields. The estimation model is learned by linear regression from structural properties of the images and known optimal configurations, that were previously obtained for low resolution data sets. Our experiments in high resolution images show that this approach is able to estimate appropriate configurations that are suitable for performing the segmentation task without requiring to re-engineer parameters. Furthermore, our combined approach reported state of the art performance on the benchmark data set HRF, as measured in terms of the F1-score and the Matthews correlation coefficient.

Open-source software platform for medical image segmentation applications.

Segmenting 2D and 3D images is a crucial and challenging problem in medical image analysis. Although several image segmentation algorithms have been proposed for different applications, no universal method currently exists. Moreover, their use is usually limited when detection of complex and multiple adjacent objects of interest is needed. In addition, the continually increasing volumes of medical imaging scans require more efficient segmentation software design and highly usable applications. In this context, we present an extension of our previous segmentation framework which allows the combination of existing explicit deformable models in an efficient and transparent way, handling simultaneously different segmentation strategies and interacting with a graphic user interface (GUI). We present the object-oriented design and the general architecture which consist of two layers: the GUI at the top layer, and the processing core filters at the bottom layer. We apply the framework for segmenting different real-case medical image scenarios on public available datasets including bladder and prostate segmentation from 2D MRI, and heart segmentation in 3D CT. Our experiments on these concrete problems show that this framework facilitates complex and multi-object segmentation goals while providing a fast prototyping open-source segmentation tool.