Active contour and deep learning methods for single-cell segmentation in microscopy images
  • Metaadatok
Tartalom: https://doktori.bibl.u-szeged.hu/id/eprint/11892/
Archívum: SZTE Doktori Értekezések Repozitórium
Gyűjtemény: Tudományterületek = Műszaki tudományok: Informatikai tudományok
Típus = Disszertáció
Cím:
Active contour and deep learning methods for single-cell segmentation in microscopy images
Létrehozó:
Tasnádi Ervin
Dátum:
2024-03-06
Téma:
01.02.01.16.02. Számítógépes látás
01.02.01.09. Bioinformatika, e-egészség, orvosi informatika
Tartalmi leírás:
This work introduces methods for single-cell segmentation of microscopy images. The developed methods are based on active contours and deep learning. In the first thesis point, a reinitialization method is developed for level sets that is based on the phase field theory. When the phase field functional is minimized, it forms a smooth transition between the two phases. This phase transition can be used as a reinitialization technique when combined with level sets. However, we show that the original phase field functional moves the zero transition away from its original position as an undesired side effect. We propose a proper combination of the original gradient based and a second order term to eliminate this effect and shown to be effective when combined with different active contour models. In the second thesis point, we propose an selective active contour model in 3D for the instance segmentation of 3D microscopy images. The proposed method uses surface and volume priors and a shape prior that is a combination of the former two to describe common shapes of biological objects. The resulting segmentation method is then embedded into a popular 3D medical image analysis software for semi-automatic segmentation shown to greatly reduce the annotation time. The third thesis point investigates the applications of the image-to-image translation method for automatic data augmentation. The first proposed method synthesizes artificial instance masks using a traditional parametric cell population simulation tool and then applies the learned image-to-image translation model to synthesize the corresponding microscopy images. The second method tries to learn the discrete instance masks directly from the data using generative adversarial networks thus enabling the synthesis of complex tissue structures. The proposed methods shown to increase the test accuracy when used with different instance segmentation models.
Nyelv:
angol
angol
angol
Típus:
Disszertáció
NonPeerReviewed
Formátum:
application/pdf
application/pdf
Azonosító:
Tasnádi Ervin Active contour and deep learning methods for single-cell segmentation in microscopy images. Doktori értekezés, Szegedi Tudományegyetem (2000-). (2024)
Kapcsolat: