The student-teacher framework guided by self-training and consistency regularization for semi-supervised medical image segmentation.
PLoS One
; 19(4): e0300039, 2024.
Article
en En
| MEDLINE
| ID: mdl-38648206
ABSTRACT
Due to the high suitability of semi-supervised learning for medical image segmentation, a plethora of valuable research has been conducted and has achieved noteworthy success in this field. However, many approaches tend to confine their focus to a singular semi-supervised framework, thereby overlooking the potential enhancements in segmentation performance offered by integrating several frameworks. In this paper, we propose a novel semi-supervised framework named Pesudo-Label Mean Teacher (PLMT), which synergizes the self-training pipeline with pseudo-labeling and consistency regularization techniques. In particular, we integrate the student-teacher structure with consistency loss into the self-training pipeline to facilitate a mutually beneficial enhancement between the two methods. This structure not only generates remarkably accurate pseudo-labels for the self-training pipeline but also furnishes additional pseudo-label supervision for the student-teacher framework. Moreover, to explore the impact of different semi-supervised losses on the segmentation performance of the PLMT framework, we introduce adaptive loss weights. The PLMT could dynamically adjust the weights of different semi-supervised losses during the training process. Extension experiments on three public datasets demonstrate that our framework achieves the best performance and outperforms the other five semi-supervised methods. The PLMT is an initial exploration of the framework that melds the self-training pipeline with consistency regularization and offers a comparatively innovative perspective in semi-supervised image segmentation.
Texto completo:
1
Colección:
01-internacional
Base de datos:
MEDLINE
Asunto principal:
Procesamiento de Imagen Asistido por Computador
Límite:
Humans
Idioma:
En
Revista:
PLoS One
Asunto de la revista:
CIENCIA
/
MEDICINA
Año:
2024
Tipo del documento:
Article
País de afiliación:
China
Pais de publicación:
Estados Unidos