Analysis and segmentation of KI-67 immunohistochemistry images for breast cancer diagnosis

For breast cancer diagnosis, the computation of the KI-67 score is a useful metric to define patient?s treatment. For a proper computation of this score, tumour and stroma areas have to be distinguished in KI-67 images. However, this is not an easy task due to the fact that in KI-67 images, tumour a...

Descripción completa

Detalles Bibliográficos
Autor: Espina i Boronat, Maria
Tipo de recurso: tesis de maestría
Fecha de publicación:2022
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/381078
Acceso en línea:https://hdl.handle.net/2117/381078
Access Level:acceso abierto
Palabra clave:Breast--Cancer--Diagnosis
Imaging systems in medicine
Immunohistochemistry
KI-67
FastCUT
U-NET
Breast Cancer
Mama--Càncer--Diagnòstic
Imatgeria mèdica
Immunohistoquímica
Àrees temàtiques de la UPC::Enginyeria de la telecomunicació::Processament del senyal::Processament de la imatge i del senyal vídeo
Descripción
Sumario:For breast cancer diagnosis, the computation of the KI-67 score is a useful metric to define patient?s treatment. For a proper computation of this score, tumour and stroma areas have to be distinguished in KI-67 images. However, this is not an easy task due to the fact that in KI-67 images, tumour and stroma regions have a similar appearance. The objective of this research is to develop a semantic segmentation model capable of separating these two regions in KI-67 images. As for training this segmentation model there is the need of ground truth masks, a large part of the project has been focused on the generation of these masks. Regarding the ground truth generation, in this research we present a way to generate KI-67 ground truth masks based on translating KI-67 images into the CK-19 domain and segment them using simple morphological operators and manual corrections. The image-to-image translation model used for the translation purpose is a FASTCUT model and the semantic segmentation model we trained is a U-Net. Moreover, we present a comparative study of the performance of the U-Net for three different fields of view scope. From this research, quite good results are obtained for both ground truth generation and semantic segmentation processes.