Breast density classification with convolutional neural networks

Pablo Fonseca, Benjamin Castañeda, Ricardo Valenzuela, Jacques Wainer

Producción científica: Capítulo del libro/informe/acta de congresoContribución a la conferenciarevisión exhaustiva

7 Citas (Scopus)


Breast Density Classification is a problem in Medical Imaging domain that aims to assign an American College of Radiology’s BIRADS category (I-IV) to a mammogram as an indication of tissue density. This is performed by radiologists in an qualitative way, and thus subject to variations from one physician to the other. In machine learning terms it is a 4-ordered-classes classification task with highly unbalance training data, as classes are not equally distributed among populations, even with variations among ethnicities. Deep Learning techniques in general became the state-of-the-art for many imaging classification tasks, however, dependent on the availability of large datasets. This is not often the case for Medical Imaging, and thus we explore Transfer Learning and Dataset Augmentationn. Results show a very high squared weighted kappa score of 0.81 (0.95 C.I. [0.77,0.85]) which is high in comparison to the 8 medical doctors that participated in the dataset labeling 0.82 (0.95 CI [0.77, 0.87]).

Idioma originalInglés
Título de la publicación alojadaProgress in Pattern Recognition, Image Analysis, Computer Vision, and Applications - 21st Iberoamerican Congress, CIARP 2016, Proceedings
EditoresCesar Beltran-Castanon, Fazel Famili, Ingela Nystrom
EditorialSpringer Verlag
Número de páginas8
ISBN (versión impresa)9783319522760
EstadoPublicada - 2017
Evento21st Iberoamerican Congress on Pattern Recognition, CIARP 2016 - Lima, Perú
Duración: 8 nov. 201611 nov. 2016

Serie de la publicación

NombreLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volumen10125 LNCS
ISSN (versión impresa)0302-9743
ISSN (versión digital)1611-3349


Conferencia21st Iberoamerican Congress on Pattern Recognition, CIARP 2016
Ciudad Lima


Profundice en los temas de investigación de 'Breast density classification with convolutional neural networks'. En conjunto forman una huella única.

Citar esto