Author: CAROLINA RETA CASTRO

Segmentación y clasificación de células con leucemia a partir de información contextual en imágenes digitales

CAROLINA RETA CASTRO (2009)

In this thesis, we propose a bone marrow cell contextual analysis methodology for

the detection of acute leukemia subtypes. The first phase of the methodology focuses

on the segmentation and identification of cellular elements from bone marrow images.

In the second phase we perform feature extraction to the cells images obtained in the

first phase and use this information to classify the cells into leukemia subtypes. This

classification can be used to diagnose patients.

The segmentation algorithm uses as contextual information the color and texture of

the image pixels to be able to separate the nucleus and cytoplasm of blood cells from

bone marrow smear images, which show heterogeneous color and texture staining and

a high cell population. The regions obtained from segmentation are later analyzed to

identify the cells in the image.

An additional algorithm to identify cells is proposed in this work. This algorithm

also uses contextual information related to the color, shape, and containment proportion

among regions to determine whether an analyzed ROI (Region of Interest) is labeled as

a probable cell, nuclei, an overlapped nuclei or cell with other image elements or decide

it is not a region of interest. If the cell identification algorithm determines that the

ROI is overlapped with other elements, it divides the ROI by using a cell separation

algorithm also proposed in this thesis. Once all of the ROIs are labeled, the cell is

identified by associating its respective nuclei and cytoplasm, which is easily obtained

by applying difference-set operations.

The evaluation of the segmentation algorithm is carried out by comparing the identified

regions with a manual segmentation. In general, an average accuracy of 95% was

achieved in nucleus and cell segmentation using real bone marrow cells images. The

accuracy is considered pretty good due to its high impact on the process of automatic

classification of acute leukemia cells subtypes.

In the cell classification phase we extract descriptive features (morphological, statistical,

texture, size ratio and eigenvalues), to the nucleus and cytoplasm. These features

were the input to several attribute selection and classification algorithms in order to

generate patterns that facilitate the identification of the type and subtype of each acute

leukemia cell in the image collection.

En este trabajo de tesis se propone una metodología de análisis contextual de células

de médula ósea para la detección de subtipos de leucemia aguda. La primera fase de la

metodología se centra en la segmentación e identificación de los elementos celulares de

imágenes de médula ósea. En la segunda fase se realiza la extracción de características

de las imágenes de células identificadas en la primera fase, para clasificarlas en subtipos

de leucemia y diagnosticar al paciente.

El algoritmo de segmentación que se propone utiliza la información contextual del

color y textura de los píxeles de la imagen para extraer el núcleo y citoplasma de

células sanguíneas en imágenes digitales de frotis de médula ósea que presentan tinciones

heterogéneas en color y textura, así como una alta población de células. Las regiones

resultantes de la segmentación se analizan posteriormente para identificar las células en

la imagen.

El algoritmo de identificación de células que se propone utiliza la información contextual

del color, la forma y la relación de contenido entre regiones para determinar

si la ROI (Región de Interés - Region of Interest) analizada es una célula, un núcleo,

probablemente sea un núcleo o una célula traslapada con otros elementos o bien la

región no es de interés. Si el algoritmo de identificación de células determina que la

ROI presenta traslapes con otros elementos, ésta es dividida mediante el algoritmo de

separación de regiones que se diseñó en esta tesis. Una vez que se identificaron todas

las ROI se asocia el núcleo a la célula para verificar que esta última realmente lo es y

se obtiene el citoplasma de la misma.

La evaluación de la segmentación de las regiones identificadas se compara con la

segmentación manual realizada por el experto. En promedio se obtuvo una precisión

del 95% en el núcleo y la célula al utilizar imágenes reales de células de médula ósea.

La precisión alcanzada en esta evaluación es considerada muy buena ya que tiene un

alto impacto en el proceso de clasificación automática de subtipos de leucemias agudas.

En el proceso de clasificación, se extrajeron características morfológicas, estadísticas,

de textura, de proporción de tamaño y valores propios del núcleo y citoplasma

para representar a las células de manera descriptiva.

Master thesis

Imaging Image segmentation Image classification Data mining CIENCIAS FÍSICO MATEMÁTICAS Y CIENCIAS DE LA TIERRA MATEMÁTICAS CIENCIA DE LOS ORDENADORES

Seguimiento de múltiples personas considerando oclusión parcial y total en escenarios estacionarios no controlados

CAROLINA RETA CASTRO (2014)

El seguimiento de múltiples personas en entornos reales es un problema desafiante,

principalmente porque la silueta deformable del cuerpo humano y la iluminación variable

del entorno cambian con el tiempo la apariencia de las personas. Esta situación

provoca una alta dificultad en la asociación temporal de la identidad de las personas.

El problema se acentúa cuando los individuos se mueven cerca de otros, se ocluyen, o

cambian abruptamente su trayectoria.

En esta tesis se propone un nuevo algoritmo de asociación temporal para el seguimiento

individual y secuencial de múltiples personas en escenarios no controlados a

partir de una cámara estacionaria. El algoritmo de asociación propuesto construye un

grafo de seguimiento a partir de un análisis de la interacción de las personas y de mediciones

con ruido proporcionadas por un esquema de detección de personas. El grafo

de seguimiento modela las relaciones espacio-temporales de las personas en la escena

para predecir y resolver oclusiones parciales y totales. Cuando se presenta un evento

de oclusión total, el algoritmo genera diversas hipótesis acerca de la ubicación de la

persona ocluida considerando 3 casos: a) la persona mantiene su misma dirección y velocidad,

b) la persona adopta la dirección y la velocidad de su oclusor, y c) la persona

permanece inmóvil durante la oclusión. Mediante el análisis del grafo de seguimiento

durante su construcción, el algoritmo propuesto es capaz de detectar falsos positivos y

falsos negativos en las mediciones de detección y también puede estimar la ubicación

de personas no detectadas u ocluidas.

Doctoral thesis

Computer vision Image processing Multiple target tracking Video surveillance Motion measurement Feature extraction CIENCIAS FÍSICO MATEMÁTICAS Y CIENCIAS DE LA TIERRA MATEMÁTICAS CIENCIA DE LOS ORDENADORES

On the implementation of a hardware architecture for an audio data hiding system

JOSE JUAN GARCIA HERNANDEZ CLAUDIA FEREGRINO URIBE RENE ARMANDO CUMPLIDO PARRA CAROLINA RETA CASTRO (2011)

Data hiding systems have emerged as a solution against the piracy problem, particularly those based on quantization have been widely used for its simplicity and high performance. Several data hiding applications, such as broadcasting monitoring and live performance watermarking, require a real-time multi-channel behavior. While Digital Signal Processors (DSP) have been used for implementing these schemes achieving real-time performance for audio signal processing, custom hardware architectures offer the possibility of fully exploiting the inherent parallelism of this type of algorithms for more demanding applications. This paper presents an efficient hardware implementation of a Rational Dither Modulation (RDM) algorithm-based data hiding system in the Modulated Complex Lapped Transform (MCLT) domain. In general terms, the proposed hardware architecture is conformed by an MCLT processor, an Inverse MCLT processor, a Coordinate Rotation Digital Computer (CORDIC) and an RDM-QIM processor. Results of implementing the proposed hardware architecture a Field Programmable Gate Array (FPGA) are presented and discussed.

Article

Data hiding Audio signal FPGA Multi-channel processing CIENCIAS FÍSICO MATEMÁTICAS Y CIENCIAS DE LA TIERRA MATEMÁTICAS CIENCIA DE LOS ORDENADORES CIENCIA DE LOS ORDENADORES

Efficient implementation of the RDM-QIM algorithm in an FPGA

JOSE JUAN GARCIA HERNANDEZ CAROLINA RETA CASTRO RENE ARMANDO CUMPLIDO PARRA CLAUDIA FEREGRINO URIBE (2009)

The RDM-QIM algorithm has been proposed as a solution to the gain attack in QIM-based data hiding schemes. Several data hiding applications, such as broadcasting monitoring and live performance watermarking, requires a real-time multi-channel behavior. While Digital Signal Processors (DSP) have been used for implementing these schemes achieving real-time performance for audio signal processing, FPGAs offer the posibility of fully exploiting the inherent parallelism of this type of algorithms for more demanding applications. This letter presents an efficient FPGA implementation of RDM-QIM algorithm that overcomes a DSP-based implementation for more than two orders of magnitude and allows real-time multi-channel behavior.

Article

Data hiding Rational dither modulation Hardware architectures FPGA CIENCIAS FÍSICO MATEMÁTICAS Y CIENCIAS DE LA TIERRA MATEMÁTICAS CIENCIA DE LOS ORDENADORES

Acute leukemia classification by ensemble particle swarm model selection

Hugo Jair Escalante Balderas Manuel Montes y Gómez Jesús Antonio González Bernal María del Pilar Gómez Gil Leopoldo Altamirano Robles CARLOS ALBERTO REYES GARCIA CAROLINA RETA CASTRO ALEJANDRO ROSALES PEREZ (2012)

Objective: Acute leukemia is a malignant disease that affects a large proportion of the world population. Different types and subtypes of acute leukemia require different treatments. In order to assign the correct treatment, a physician must identify the leukemia type or subtype. Advanced and precise methods are available for identifying leukemia types, but they are very expensive and not available in most hospitals in developing countries. Thus, alternative methods have been proposed. An option explored in this paper is based on the morphological properties of bone marrow images, where features are extracted from medical images and standard machine learning techniques are used to build leukemia type classifiers. Methods and materials: This paper studies the use of ensemble particle swarm model selection (EPSMS), which is an automated tool for the selection of classification models, in the context of acute leukemia classification. EPSMS is the application of particle swarm optimization to the exploration of the search space of ensembles that can be formed by heterogeneous classification models in a machine learning toolbox. EPSMS does not require prior domain knowledge and it is able to select highly accurate classification models without user intervention. Furthermore, specific models can be used for different classification tasks. Results: We report experimental results for acute leukemia classification with real data and show that EPSMS outperformed the best results obtained using manually designed classifiers with the same data. The highest performance using EPSMS was of 97.68% for two-type classification problems and of 94.21% for more than two types problems. To the best of our knowledge, these are the best results reported for this data set. Compared with previous studies, these improvements were consistent among different type/subtype classification tasks, different features extracted from images, and different feature extraction regions. The performance improvements were statistically significant.Weimproved previous results by an average of 6% and there are improvements of more than 20% with some settings. In addition to the performance improvements, we demonstrated that no manual effort was required during acute leukemia type/subtype classification.

Conclusions: Morphological classification of acute leukemia usingEPSMSprovides an alternative to expensive diagnostic methods in developing countries. EPSMS is a highly effective method for the automated construction of ensemble classifiers for acute leukemia classification, which requires no significant user intervention. EPSMS could also be used to address other medical classification tasks.

Article

Ensemble learning Swarm optimization Full model selection Morphological classification Analysis of bone marrow cell images Acute leukemia classification CIENCIAS FÍSICO MATEMÁTICAS Y CIENCIAS DE LA TIERRA MATEMÁTICAS CIENCIA DE LOS ORDENADORES CIENCIA DE LOS ORDENADORES