Accurate Whole Cell Instance Segmentation from Brightfield Images

Prytula, Yaroslav

Home
→
Students Research & Project Works | Роботи студентів
→
Факультет прикладних наук
→
Бакалаврська програма "Комп'ютерні науки"
→
2023
→
View Item

dc.contributor.author	Prytula, Yaroslav
dc.date.accessioned	2024-02-14T09:07:17Z
dc.date.available	2024-02-14T09:07:17Z
dc.date.issued	2023
dc.identifier.citation	Prytula, Yaroslav. Accurate Whole Cell Instance Segmentation from Brightfield Images / Yaroslav Prytula; Supervisor: Dmytro Fishman; Ukrainian Catholic University, Department of Computer Sciences. – Lviv: 2023. – 57 p.: ill.
dc.identifier.uri	https://er.ucu.edu.ua/handle/1/4406
dc.language.iso	en	uk
dc.title	Accurate Whole Cell Instance Segmentation from Brightfield Images	uk
dc.type	Preprint	uk
dc.status	Публікується вперше	uk
dc.description.abstracten	Semantic and instance segmentations have revolutionized biomedical image anal- ysis, playing a crucial role in numerous biological applications. The development of accurate segmentation pipelines has enabled fast and reliable image analysis. Pre- vious state-of-the-art methods in cellular biology rely on accurate cell segmentation without preserving knowledge of overlapping instances. In this work, we first show that extending the model by introducing multiple decoupled decoders for multi-task learning greatly helps in scene understanding and results in high-fidelity segmen- tations. Furthermore, we identify cases of overlap occurrence and construct prob- ability maps based on cell spatial proximity. Additionally, to overcome the lack of annotated samples, we introduce a way to synthesize brightfield images and show that applying overlap-aware weight maps directly to the loss function guides the model to attend to regions of occluded cells, thus improving segmentation perfor- mance. We then propose an approach to extend our model to perform instance seg- mentation. Compared to previous state-of-the-art approaches, we utilize a concep- tually novel method of learning instance activation maps that highlight informa- tive regions for different cells for global awareness. Without bells and whistles, we combine multi-task learning with overlap awareness for instance segmentation, and show that our approach achieves state-of-the-art results.	uk