Comparing modern segmentation architectures under low data regime for PET-CT tumor segmentation - LATIM - IMT Atlantique Accéder directement au contenu
Communication Dans Un Congrès Année : 2022

Comparing modern segmentation architectures under low data regime for PET-CT tumor segmentation

Résumé

In medical imaging, low data regimes arise from practical situations where not only data labeling but also data collection itself is expensive. Most modern segmentation architectures compensate data scarcity by the intensive use of data augmentation. However, another important lever for improved generalization power is the neural architecture itself. In this context, the goal of this paper is to provide insights into the performances of various modern hybrid and Transformers-based segmentation networks under various data regimes. We believe this comparative analysis is crucial to better understand and guide future research on segmentation scenarios suffering from data scarcity. We conducted a series of controlled studies to analyze PET-CT head and neck tumor segmentation performance of hybrid models versus pure Transformer-based models with a variable number of training subjects. One path versus multipath encoders were also investigated to study the management of multi-modality. Results on the publicly-available HECKTOR 2021 dataset show that hybrid pipelines generally tend to outperform simple Transformers-based models and perform similarly with respect to hierarchical Transformers. In these experiments, multipath encoders surpass the one path strategy, and no particular improvement using multi-modal Transformers was observed.
Fichier principal
Vignette du fichier
IEEE_MIC_2022___Comparing_modern_segmentation_architectures_under_low_data_regime_for_PET_CT_tumor_segmentation.pdf (341.26 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04040739 , version 1 (22-03-2023)

Identifiants

  • HAL Id : hal-04040739 , version 1

Citer

Gustavo Andrade-Miranda, Vincent Jaouen, Dimitris Visvikis, Pierre-Henri Conze. Comparing modern segmentation architectures under low data regime for PET-CT tumor segmentation. NSM MIC 2022: IEEE Nuclear science symposium and medical imaging conference, Nov 2022, Milan, Italy. ⟨hal-04040739⟩
25 Consultations
26 Téléchargements

Partager

Gmail Facebook X LinkedIn More