Work

Siirto-oppiminen histopatologisessa kuva-analyysissä

Year: 2023
Type: dissertation
Abstract: A histopathological analysis is the main diagnostic tool for many diseases, and because of the aging population, the shortage of pathologists has been a concern in many countries. The use of deep learning has become popular in the field of histopathology, as it can speed up the diagnosis process and decrease the workload of pathologists. Histopathological data has typically a lot of variation due to differing practices and tools used in laboratories and hospitals, making the development process of neural networks often time consuming and expensive. Thus better methods for developing neural networks are needed. Neural networks trained with histopathological images generalize poorly for images from different scanners. Thus, adding new data to a project usually requires adding more training data, and training a new AI model. Transfer learning is a machine learning method, that can be used to speed up the training of the model and decrease the required amount of data. This thesis aimed to find out, how and when transfer learning should be used in the training of neural networks for histopathological image analysis. The goal is to speed up the development process in a situation where data from different scanners is added to a project. Prostate cancer microarray images with epithelium annotations were used for the study. Transfer learning does not always improve the model performance, but it speeds up the training of a neural network. The smaller the training data set is, the more transfer learning can improve the model performance. The results were similar both with overlapping and not overlapping source and target data sets. Consecutive use of transfer learning, "chaining", did not have a negative effect on the model performance either.
Author: Tähti, Linn
Language: English
Open Access status: green
OpenAlex ID: W7113077828
Cites: 34
Cited by:
Related to: 10