Repositori institucional URV
Español Català English
Belongs to TFM:SerieGeneralMESIIA collection
TITLE:
Segmentation and classification of breast cancer pathologies in histological images based on morphological patterns - TFM:357
Handle: https://hdl.handle.net/20.500.11797/TFM357

Student:Manzi, Berardo Mario
Language:Anglès
Title in original language:Segmentation and classification of breast cancer pathologies in histological images based on morphological patterns
Title in different languages:Segmentation and classification of breast cancer pathologies in histological images based on morphological patterns
Keywords:Segmentation, Breast Cancer, Artificial Intelligence
Subject:Enginyeria informàtica
Abstract:Mammography screening for breast cancer detection is a routine control for women aged 40 or older. Eventual suspects of a tumor might lead to biopsies, to allow for histological studies. The resulting images will be analyzed by expert pathologist to detect any of the common cancer types (adenosis, fibroadenoma, phyllodes tumor, tubular adenoma, ductal carcinoma, lobular carcinoma, mucinous carcinoma, papillary carcinoma), to define a proper treatment. In our work, we aim to develop an algorithm capable of detecting such cancer types, which could support medical personnel to emit a diagnosis. For this purpose, we exploited machine learning techniques, training convolutional neural networks to identify the various cancer classes through a two stages approach. In the first stage, the networks are expected to segment the images by individuating the regions of relevance for diagnosis, such as the connective tissue (or stroma), and epithelium cells, whose alterations are signs of possible disease. The second stage focuses on classifying the outputs of the first stage into the eight types of cancers, or, at least, into benign and malignant tumors. This approach attempts to get some insight into the meaning of the features learned by a neural network, in contrast with the usual end-to-end classification methods commonly employed, which are capable to work properly but whose inner mechanism is unknown even to the designer, acting, thus, as a black-box. We compare our approach to such end-to-end classification scheme and show that we are capable to obtain similar accuracies while, at the same time, giving some insight into the relevant visual features.
Project director:Romaní-Also, Santiago
Department:Enginyeria Informàtica i Matemàtiques
Education area(s):Enginyeria de la Seguretat Informàtica i Intel·ligència Artificial
Entity:Universitat Rovira i Virgili (URV)
TFM credits:9
Work's public defense date:2018-09-12
Academic year:2017-2018
Confidenciality:No
Subject areas:Computer engineering
Creation date in repository:2018-02-11
Search your record at:

Available files
FileDescriptionFormat
MemòriaMemoryapplication/pdf

Information

© 2011 Universitat Rovira i Virgili