Oscar Perdomo

Training Deep Convolutional Neural Networks with Active Learning for Exudate Classification in Eye Fundus Images.

Training deep convolutional neural network for classification in medical tasks is often difficult due to the lack of annotated data samples. Deep convolutional networks (CNN) has been successfully used as an automatic detection tool to support the grading of diabetic retinopathy and macular edema. Nevertheless, the manual annotation of exudates in eye fundus images used to classify the grade of the DR is very time consuming and repetitive for clinical personnel. Active learning algorithms seek to reduce the labeling effort in training machine learning models. This work presents a label-efficient CNN model using the expected gradient length, an active learning algorithm to select the most informative patches and images, converging earlier and to a better local optimum than the usual SGD (Stochastic Gradient Descent) strategy. Our method also generates useful masks for prediction and segments regions of interest.

Morphological analysis of T-wave in vectorcardiographic leads system by a bi-Gaussian approach in patients under effect of salbutamol

There are several models of decomposition of the electrocardiogram (ECG). Some of these models are intended to describe the ECG signal, and others are more specific to extract the relevant information relating to individual waveform which contributes to explain the P-QRS complex. The latter approach may be particularly suitable for a portion where a morphological analysis of the ECG is of particular interest, as the cardiac repolarization segment or T-wave. This study aims: to model and detect useful patterns in the evaluation of T wave morphology, which explains the different changes in ventricular repolarization during inhalation of Salbutamol.

A Method to Detect Potentially Malignant Skin Lesions Through Image Segmentation

Melanoma is a form of skin cancer responsible for most cancer deaths. The detection of melanoma in an early stage is still a challenge with high importance. The aim of this paper was the design of a system to analyse the mole’s characteristics, based on the ABCDE criteria, for identifying potentially malignant skin lesions. The segmentation was based on Otzu thresholding, chain code and skeletonization methods were used to compute the main features: asymmetry, border, color and texture, for a future classification of skin lesions as normal or potentially malignant melanoma. The system was evaluated successfully using 92 images with 27 benign and 65 malignant melanomas from Dermatology Information System (DermIS).

Glaucoma Diagnosis from Eye Fundus Images Based on Deep Morphometric Feature Estimation

Glaucoma is an ophthalmic disease related to damage in the optic nerve and it is without symptoms in its early stages. Left untreated, it can lead to vision limitation and blindness. Eye fundus images have been widely accepted by medical personnel to examine the morphology and texture of the optic nerve head and the physiologic cup but glaucoma diagnosis is still subjective and without clear consensus among experts. This paper presents a multi-stage deep learning model for glaucoma diagnosis based on a curriculum learning strategy. In curriculum learning, a model is sequentially trained to solve incrementally difficult tasks. Our proposed model includes the following stages: segmentation of the optic disc and physiological cup, prediction of morphometric features from segmentations, and prediction of disease level (healthy, suspicious and glaucoma). The experimental evaluation shows that our proposed method outperforms conventional convolutional deep learning models from the state of the art reported on the RIM-ONE-v1 and DRISHTI-GS1 datasets with an accuracy of 89.4% and an AUC of 0.82 respectively.

Convolutional network to detect exudates in eye fundus images of diabetic subjects

Diabetic retinopathy has several clinical data sources for medical diagnosis, but the lack of tools to process the data generates a subjective and unclear diagnosis. The use of convolutional networks to analyze and extract features in eye fundus images may help with an automatic detection to support medical personnel in the grading of diabetic retinopathy. This paper presents a description of convolutional neural networks as a good methodology to detect and discriminate between exudate and healthy regions in eye fundus images.

Simultaneous Measurement of Trunk Orientation and Centre of Pressure for Postural Stability Evaluation

Human balance control system allows people to independently perform acts of daily living and to avoid falls that can cause injuries and hospitalization. Aging or various pathologies may cause disorders in the balance system and result in wide variety of problems and generally decreased quality of life. Researchers and health professionals therefore strive to develop effective methods and instruments for prompt diagnosis and rehabilitation of balance disorders. A common approach used to detect balance problems in scientific studies and clinical applications is to analyze various parameters of postural stability gathered during quiet stance. This paper presents an experimental setting for postural stability evaluation composed of two independent systems. A commercial force platform (Ecowalk) and a sensor node with multiple inertial and magnetic sensors are combined to simultaneously measure trunk orientation and center-of-pressure (COP) during quiet stance. A preliminary study with 8 healthy subjects was conducted in order to test the system functionality. Trunk orientation and COP were measured in four different conditions for every subject as they tried to maintain steady upright position. The area of trunk sway trajectory and COP variability were calculated from orientation and force platform measurements. The goal of this study was to explore whether use of the created experimental setting can provide a measure of two different biomechanical properties related to the balance control system during quiet stance trials. The initial results are promising as they have shown consistency in both variables during trials in different conditions, but further measurements with more subjects are needed for stronger conclusions.

Modeling and simulation of ventricular repolarization or T-wave of ECG of subjects with type 1 diabetes during adrenaline infusions

Increased catecholamine activity, both endogenous, and through medications, it has long been recognized to increase the risk of cardiac death. This paper proposes the use of the Bueno-Orovio - Cherry - Fenton model to understand the behavior of ventricular cells: Endocardium, Myocardium and Epicardial, through the transmural ECG generated by the weighted sum of action potentials of the three ventricular cells. These results were validated with a dataset of Vectorcardiograms according to Frank VCG-lead system (X, Y, Z) of 22 subjects with type 1 diabetes along adrenaline infusion, aged 20 to 40 years. Moreover, it was analyzed all 12-lead ECG, calculated from VCGs by using the transform Dower. Finally, due to the sensitivity of the potassium current in the myocardium cells and the simulated transmural ECG, decrements considerably bigger than those used in this study could generate early afterdepolarization (EAD), flattening and inversion of the T-wave.

Pilot study: Assessing repeatability of the EcoWalk platform resistive pressure sensors to measure plantar pressure during barefoot standing

Plantar pressure provides useful information to assess the feets condition. These systems have emerged as popular tools in clinical environment. These systems present errors and no compensation information is presented by the manufacturer, leading to uncertainty in the measurements. Ten healthy subjects, 5 females and 5 males, were recruited. Lateral load distribution, antero-posterior load distribution, average pressure, contact area, and force were recorded. The aims of this study were to assess repeatability of the EcoWalk system and identify the range of pressure values observed in the normal foot. The coefficient of repeatability was less than 4% for all parameters considered.