1st ITMA

Abstract information

English title:	Region of interest classifier in mammographic images using logistic regression
Spanish title:	Clasificador de regiones de inter�s en im�genes mamogr�ficas usando regresi�n log�stica
English abstract:	The early detection of breast cancer is important for providing opportune treatment and improve the possibilities of survival of the patient. For this reason, screening techniques are implemented, such as full-field digital mammography, allowing the radiologist to give a more precise diagnosis. Recently, machine learning has shown great potential for computer-aided diagnosis. It has been shown that these methodologies reduce the incidence of false positives since they are more accurate in the classification of the regions of interest. In this work, we propose a model for the classification of lesions in mammograms based on logistic regression where the predictor variables are features extracted by the software Openbreast. For the implementation of the classifier, we use a public dataset named INbreast which contains 410 images and includes cases of women with both breasts affected and different lesions such as masses, calcifications, asymmetries, and distortions. It also contains labels and annotations made by radiologists where regions of interest were previously selected. For training and testing the performance of the classifier, different regions of interest were selected and cropped for obtaining two subsets of 106 images each one, the first containing normal images and the other abnormal images. After this, texture and shape features of the regions are extracted using the Openbreast software. Texture features include different methods such as Gray-Level Co-Occurrence Matrix (GLCM), Gray-Level Run-Length Matrix (GLRL), Gray-Level Histogram Analysis (GLHA), and Gray-Level Sharpness Measurement. Additionally, the shape features were extracted based on mathematical descriptors of the region such as area, concavity, circularity, and perimeter. In total, 44 features were extracted, corresponding to 34 and 10 texture and shape features, respectively. The proposed method was validated using 5-folds cross-validation. Performance was measured in terms of classification accuracy and are under the ROC curve (AUC), obtaining 73.1% and 0.78, respectively. In order to improve the results, further work on the exploration of extraction and selection of features is warranted.
Spanish abstract:	La detecci�n temprana del c�ncer de seno es importante para proporcionar un tratamiento oportuno y as�, mejorar las posibilidades de supervivencia de la paciente. Debido a esto, se implementan t�cnicas de cribado, tales como la mamograf�a digital de campo completo, permiti�ndole al radi�logo proporcionar un diagn�stico m�s preciso. Recientemente, las t�cnicas de Machine Learning han demostrado un gran potencial para el diagn�stico asistido por computadora. Se ha demostrado que estas metodolog�as reducen la incidencia de falsos positivos, ya que son m�s precisas en la clasificaci�n de regiones de inter�s. En este trabajo, se propone un modelo para la clasificaci�n de lesiones en mamograf�as basado en regresi�n log�stica, donde las variables predictoras son caracter�sticas extra�das por el software Openbreast. Para la implementaci�n del clasificador se us� una base de datos p�blica llamada INbreast, la cual contiene 410 im�genes e incluye casos de mujeres con ambos senos afectados y diferentes lesiones, tales como masas, calcificaciones, asimetr�as y distorsiones. Esta tambi�n contiene etiquetas y anotaciones hechas por radi�logos, donde las regiones de inter�s fueron seleccionadas previamente. Para entrenar y evaluar el desempe�o del clasificador, se seleccionaron diferentes regiones de inter�s y se recortaron de modo que se obtuvieran dos subconjuntos de 106 im�genes cada uno, el primero contiene im�genes normales y el otro im�genes anormales. Despu�s de esto, se extraen caracter�sticas de textura y forma usando el software Openbreast. Para la extracci�n de caracter�sticas de textura se emplearon m�todos como son la matriz de coocurrencia de nivel de gris, matriz de longitud de ejecuci�n de nivel de gris, an�lisis de histograma de nivel de gris y medida de nitidez de nivel de gris. Adicionalmente, las caracter�sticas de forma se extrajeron a partir de descriptores matem�ticos de la regi�n, tales como �rea, concavidad, circularidad y per�metro. En total, se extrajeron 44 caracter�sticas, de las cuales 34 corresponden a textura y las 10 restantes a caracter�sticas de forma. El m�todo propuesto fue validado usando validaci�n cruzada de 5 folds. El desempe�o fue medido en t�rminos de precisi�n de clasificaci�n y �rea bajo la curva ROC, obteni�ndose 73.1% y 0.78, respectivamente. Para mejorar los resultados se requiere seguir trabajando en la exploraci�n de maneras para la extracci�n y selecci�n de caracter�sticas.
Topic:
Presentation Type:
Speaker:

Authors

Name	Institution	Email
Natalia Cabeza	Universidad Industrial de Santander	nataliajohanacabeza@gmail.com
Maria Angelica Bravo	Universidad Industrial de Santander	maria2182344@correo.uis.edu.co

Presentation

Room:	Aula Máxima de Ingeniería Mecánica Auditorium
Date:	1 December 2022
Hour:	09:00 - 09:20

Home

1^st ITMA

Innovative Technologies for Medical Applications

Universidad Industrial de Santander

Contact information:

Research group CPS: cps.info@uis.edu.co

Copyright © FORISTOM Foundation, 2022
All rights reserved