Designing Convolutional Neural Networks for Classification of Metastatic Tissue in the Lymph Nodes : Evaluation of the Effect of Feature Map Distributions Across Convolutional Layers on Network Performance
Year: 2022
Type: dissertation
Abstract: Cancer is one of the leading causes of mortality in the world. It is estimated that about 20% of males and 17% of females will suffer it at some point in their lives, while 13% of males and 9% of females will die from it [1]. Cancer is caused by abnormal cell growth which can spread through the body in a process called metastasis. One of the ways to diagnose and evaluate the severity of cancer is to analyse the lymph nodes for presence of metastatic tissue. Recently, Artificial Neural Networks (ANNs) have been used to aid the detection of such tissue in images, but there are many design decisions to consider in making a network for this task. One of the most important design decisions when developing an ANN is the choice of the network, as this will define how the model will evaluate patterns from the input. Convolutional Neural Networks are the main ANN architecture used for image analysis, as these use filters to evaluate how prevalent different shapes or patterns are in an image. My research has studied how different ways of using these filters affects lymph node image classification results. From the worst to best performing models tested, there is a 0.0739 accuracy gap (from 0.8494 to 0.9234), representing an 8.709% increase in performance. Within models of the same size in terms of number of patterns analysed, there are gaps of up to 0.035 accuracy. My research has found network size to be closely related to training time, although different designs have different training complexities, so a case can be made for models of a same size that perform worse but also train faster. The main conclusion of this project is that the way features are extracted from lymph node images in a Neural Network clearly affect its performance, but also its training complexity. With this knowledge, one can test different approaches and designs according to their requirements and resources.
Author: Ginebra Solanellas, Pere
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7027884309