Multi-Task, Multi-Domain Deep Segmentation with Shared Representations\n and Contrastive Regularization for Sparse Pediatric Datasets

Abstract

Automatic segmentation of magnetic resonance (MR) images is crucial for\nmorphological evaluation of the pediatric musculoskeletal system in clinical\npractice. However, the accuracy and generalization performance of individual\nsegmentation models are limited due to the restricted amount of annotated\npediatric data. Hence, we propose to train a segmentation model on multiple\ndatasets, arising from different parts of the anatomy, in a multi-task and\nmulti-domain learning framework. This approach allows to overcome the inherent\nscarcity of pediatric data while benefiting from a more robust shared\nrepresentation. The proposed segmentation network comprises shared\nconvolutional filters, domain-specific batch normalization parameters that\ncompute the respective dataset statistics and a domain-specific segmentation\nlayer. Furthermore, a supervised contrastive regularization is integrated to\nfurther improve generalization capabilities, by promoting intra-domain\nsimilarity and impose inter-domain margins in embedded space. We evaluate our\ncontributions on two pediatric imaging datasets of the ankle and shoulder\njoints for bone segmentation. Results demonstrate that the proposed model\noutperforms state-of-the-art approaches.\n

Ähnliche Arbeiten