Few-Shot Airway-Tree Modeling using Data-Driven Sparse Priors

Abstract

The lack of large annotated datasets in medical imaging is an intrinsic burden for supervised Deep Learning (DL) segmentation models. Few-shot learning approaches are cost-effective solutions to transfer pre-trained models using only limited annotated data. However, such methods can be prone to overfitting due to limited data diversity especially when segmenting complex, diverse, and sparse tubular structures like airways. Furthermore, crafting informative image representations has played a crucial role in medical imaging, enabling discriminative enhancement of anatomical details. In this paper, we initially train a data-driven sparsification module to enhance airways efficiently in lung CT scans. We then incorporate these sparse representations in a standard supervised segmentation pipeline as a pretraining step to enhance the performance of the DL models. Results presented on the ATM public challenge cohort show the effectiveness of using sparse priors in pre-training, leading to segmentation Dice score increase by 1% to 10% in full-scale and few-shot learning scenarios, respectively.