University of Twente Student Theses
Mammography segmentation using visual prompts and few samples
Bronkhorst, Matteo (2024) Mammography segmentation using visual prompts and few samples.
|
PDF
1MB |
| Abstract: | To improve the effectiveness of mammography screening, developing better diagnostic tools is paramount. For large populations, automated annotation of patient exams can significantly alleviate the workload of radiologists. Segmentation of tumorous regions in an image provides more localization context than classification, but the lack of fine-grained segmentation labels often complicates training segmentation models directly on the data available at local hospitals. We optimize three state-of-the-art segmentation models for mammography using the CBIS-DDSM, and evaluate their performance on a manually annotated subset of our private dataset. Out of a U-Net, Segmentation Transformer, and Segment Anything Model, the latter performed best by a large margin on the CBIS-DDSM, achieving an IoU of 27.42% as opposed to the U-Net achieving 9.11% and the Segmentation Transformer with 8.26%. However, performance on this public mammography dataset was unrepresentative of the zero-shot transfer performance on our private dataset. Future research should focus on similarly assessing the usability of other public mammography datasets for training diagnostic tools that are effective in clinical settings. |
| Item Type: | Essay (Master) |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 44 medicine, 54 computer science |
| Programme: | Computer Science MSc (60300) |
| Link to this item: | https://purl.utwente.nl/essays/99922 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page