Classification of Gastrointestinal Lesions from Endoscopy Data

Date

2026-8-6

Author

Sönmez, Necdet Can

Metadata

Show full item record

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

Item Usage Stats

40
views

11
downloads

Cite This

Accurate classification of gastrointestinal endoscopy images is important for diagnosis, yet manual assessment is subjective and subtle lesions may be overlooked. This thesis investigates endoscopic image classification using a vision transformer that was pretrained on large scale endoscopic data as a frozen feature extractor. A lightweight classification head is trained for each of six public datasets which are Kvasir V1, Kvasir V2, Kvasir Capsule, HyperKvasir, GastroVision and WCEBleedGen. Focal loss is used to reduce the effect of class imbalance and of easily classified samples. All datasets are evaluated under the same 5 fold stratified cross validation protocol. In addition, a reduced global label space of 34 classes is defined and a stacking ensemble is trained on the concatenated outputs of five of these heads under a nested cross validation scheme. The ensemble improves the results for the two datasets that contain similar images and not for the remaining three.

Subject Keywords

Computer Vision, Endoscopy, Image Classification

URI

https://hdl.handle.net/11511/120091

Collections

Graduate School of Natural and Applied Sciences, Thesis