Date

2022-9-2

Author

İnci, Alperen

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

372
views

253
downloads

Cite This

Many people suffer from Ulcerative Colitis (UC), which is a chronic inflammatory bowel disease. UC exhibits itself as ulcers, inflammation, and sores in the colon. In order to provide proper treatment, an expert physician needs to assess the severity of the disease. However, physicians have disagreements on the severity of the disease in terms of the widely used Mayo scoring. This brings in the need for reliable and au- tomated methods. Even though automated disease diagnosis using medical imaging has become a trending topic, labeling data in the medical field requires the agreement of a committee of experts and this is a limiting factor in obtaining a sufficient amount and variety of data for robust model training. Although generative data augmenta- tion methods have been proven to increase data diversity and quality, directly adding synthetic samples may not contribute to a model’s performance, and may even result in a performance drop. Some of these generated samples might be unrealistic, or already have been learned by the model and predicted with a high confidence score. In this thesis, we propose generation of a synthetic data pool, a subset of which is then selected and used for training. The data pool is created by generating samples with different truncation parameters. Then, active learning approaches are adopted to select informative synthetic samples among these samples in the data pool for model training. The results show that the results are favorable, and the performance is more stable compared to the baseline method and random selection of generated samples.

Subject Keywords

active learning, dataset cleaning, image classification, synthetic sample selection

URI

https://hdl.handle.net/11511/98788

Collections

Graduate School of Informatics, Thesis