Immunohistochemistry images analysis with convolutional and graph neural networks
Breast cancer is one of the most prevalent diseases in our society. The survival rate is increasing thanks to medical advances and the early detection of the disease through various diagnostic tests. To achieve this, a tissue sample is extracted from a biopsy, stained with different immunohistochemi...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2024 |
| 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:2117/406217 |
| Acceso en línea: | https://hdl.handle.net/2117/406217 |
| Access Level: | acceso abierto |
| Palabra clave: | Neural networks (Computer science) Artificial intelligence Histologia càncer de mama xarxes neuronals convolucionals xarxes neuronals basades en grafs Histology Breast cancer Convolutional Neural Networks Graph Neural Networks Xarxes neuronals (Informàtica) Intel·ligència artificial Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial |
| Sumario: | Breast cancer is one of the most prevalent diseases in our society. The survival rate is increasing thanks to medical advances and the early detection of the disease through various diagnostic tests. To achieve this, a tissue sample is extracted from a biopsy, stained with different immunohistochemical techniques and analysed to quantify the tumour cells. For instance, the Estrogen Receptor (ER), Progesterone Receptor (PR), and Ki67 are some of the most commonly used stains to identify tumour grade. This work is developed within the DigiPatICS project, which aims to provide algorithms for the Institut Català de la Salut (ICS). Our objective is to develop a method to segment and classify cells present in breast tissue to obtain meaningful indicators that will support pathologists in the diagnosis of tumour. The proposed method uses a Convolutional Neural Network (specifically, the Hover-Net) followed by a Graph Neural Network (GNN) to analyse these immunohistochemical stains. In the first stage, the cells are segmented and classified, and then the resulting prediction is used to construct a graph. This information is then processed by the GNN to obtain the final classification of the identified instances. The results indicate that the proposed method helps to classify the cells in the three analysed stains: ER, PR, and Ki67. We studied the effect of several factors of the method, including the Hover-Net and the Graph Neural Network, as well as the features used to represent each segmented instance and in the conditions to generate the graph. In addition, we have interpreted the results and identified significant aspects, such as the influence of certain nodes in classifying their neighbours in some situations. This information is provided by the attention weights assigned to each pair of neighbours by certain layers used in the GNN. |
|---|