Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Ali Keshavarzi, Elsa Angelini

TL;DR

This paper initially train a data-driven sparsification module to enhance airways efficiently in lung CT scans and incorporates these sparse representations in a standard supervised segmentation pipeline as a pretraining step to enhance the performance of the DL models.

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.

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

TL;DR

This paper initially train a data-driven sparsification module to enhance airways efficiently in lung CT scans and incorporates these sparse representations in a standard supervised segmentation pipeline as a pretraining step to enhance the performance of the DL models.

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.
Paper Structure (11 sections, 3 equations, 3 figures, 1 table)

This paper contains 11 sections, 3 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Convolutional Sparse Coding
  4. Segmentation Network Architectures
  5. Pretraining and Few-shot learning
  6. Experiments and Results
  7. Training and Implementation Details
  8. Dataset and Preprocessing
  9. Segmentation Results
  10. Conclusion
  11. acknowledgements

Figures (3)

  • Figure 1: CDL and CSC pipeline for creating sparse representations on lung CT scans.
  • Figure 2: Comparison between CSC encoding, MIP, and Vesselness filter on one axial and one coronal slice.
  • Figure 3: 3D reconstruction of segmented airway trees on 3 cases: GT (green) and predictions (red); First row: results from full training on the whole training cohort ATM22 (baseline). Second and third rows: results from few-shot (FS) learning on original CTs in MiniATM22 (2nd row) and with pre-training on $\text{MiniATM22}_{sp}$ (3rd row).