Loading...
Developing Active Learning Methods to Improve Classification of Medical Images
Najafi, Mostafa | 2023
60
Viewed
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 56331 (19)
- University: Sharif University of Technology
- Department: Computer Engineering
- Advisor(s): Sharifi Zarchi, Ali
- Abstract:
- With the growing use of machine learning algorithms, especially in deep neural networks, the need for annotated data for supervised learning has also increased. In many cases, it is possible to collect data widely, but annotating all of these data is usually very time-consuming, expensive, and even impossible in some cases. The goal of active learning algorithms is to maximize the model’s performance with the least annotated data. Active learning algorithms are iterative algorithms that train the model in each iteration with the current annotated data. Then, using the results of the model on the remaining data without annotation, select some new data to annotate. This process usually continues until reaching a specific criterion, such as the performance of the model on the test data set or the end of the annotation budget. In this research, we present some improvements on one of the active learning methods for image classification. For this purpose, self-supervised learning has been used to pre-training the model and selecting the initial points, and the focal loss function has been used to reduce the problem of data imbalance during training. Finally, these new improvements and the basic method have been tested on some image classification data sets, one of which is related to medical images, to compare their performance
- Keywords:
- Active Learning ; Medical Images ; Deep Neural Networks ; Self-Supervised Learning ; Annotation