Non-Conditional Anatomically-Accurate 2D Synthetic Mask Generation
Fuente:
Zenodo
Salvato in:
| Autori principali: | , , , |
|---|---|
| Natura: | Recurso digital |
| Lingua: | inglese |
| Pubblicazione: |
Zenodo
2025
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866902340314857472 |
|---|---|
| author | Azevedo, Diogo Amaral Sousa, Pedro Pereira, Tânia Oliveira, Helder |
| author_facet | Azevedo, Diogo Amaral Sousa, Pedro Pereira, Tânia Oliveira, Helder |
| contents | <p>Although the use of AI models in medicine reveals great potential, the use of medical images for the training of models understandingly raises ethical and privacy concerns. This study aims to implement a WGAN-GP model that uses a set of lung CT scans for cancer-suffering patients to generate accurate 2D synthetic semantic segmentation masks, by segmenting each CT scan into semantic masks. To compare model’s performance, different sample resolutions and hyperparameters were experimented with. Results obtained demonstrate the model’s capability to correctly map lung anatomy and segment its different components, thus producing realistic and feasible semantic segmentation masks. While current findings are limited to 2D and sensitive to sample resolution, prospects envision the branching out into 3D medical-grade and more complex samples. Said results highlight the potential for such architectures to be used in tandem with mask-conditioned generative models and two-step data augmentation.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_17247301 |
| institution | Zenodo |
| language | eng |
| publishDate | 2025 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Non-Conditional Anatomically-Accurate 2D Synthetic Mask Generation Azevedo, Diogo Amaral Sousa, Pedro Pereira, Tânia Oliveira, Helder Generative AI Deep Learning Mask Synthesis Synthetic Semantic Segmentation Lung cancer <p>Although the use of AI models in medicine reveals great potential, the use of medical images for the training of models understandingly raises ethical and privacy concerns. This study aims to implement a WGAN-GP model that uses a set of lung CT scans for cancer-suffering patients to generate accurate 2D synthetic semantic segmentation masks, by segmenting each CT scan into semantic masks. To compare model’s performance, different sample resolutions and hyperparameters were experimented with. Results obtained demonstrate the model’s capability to correctly map lung anatomy and segment its different components, thus producing realistic and feasible semantic segmentation masks. While current findings are limited to 2D and sensitive to sample resolution, prospects envision the branching out into 3D medical-grade and more complex samples. Said results highlight the potential for such architectures to be used in tandem with mask-conditioned generative models and two-step data augmentation.</p> |
| title | Non-Conditional Anatomically-Accurate 2D Synthetic Mask Generation |
| topic | Generative AI Deep Learning Mask Synthesis Synthetic Semantic Segmentation Lung cancer |
| url | https://doi.org/10.5281/zenodo.17247301 |