Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CT evaluation of 2D and 3D holistic deep learning methods for the volumetric segmentation of airway lesions

Amel Imene Hadj Bouzid, Baudouin Denis de Senneville, Fabien Baldacci, Pascal Desbarats, Patrick Berger, Ilyes Benlala, Gaël Dournes

TL;DR

This study compares 2D and 3D holistic segmentation of cystic fibrosis airway lesions on CT using nnU-Net, introducing a loss tailored to better segment small structures. The 3D model consistently outperforms the 2D counterpart on complex lesions like mucus plugs and consolidations, though a loss-augmented 2D model shows notable improvements without surpassing the 3D results. External validation against pulmonary function tests reveals clinically meaningful correlations for certain lesion volumes, supporting robustness across architectures. Interpretability analyses (Grad-CAM) and Bayesian uncertainty estimation indicate reliable localization and comparable prediction confidence, underscoring the potential of fully automated volumetric scoring for CF severity in clinical practice.

Abstract

This research embarked on a comparative exploration of the holistic segmentation capabilities of Convolutional Neural Networks (CNNs) in both 2D and 3D formats, focusing on cystic fibrosis (CF) lesions. The study utilized data from two CF reference centers, covering five major CF structural changes. Initially, it compared the 2D and 3D models, highlighting the 3D model's superior capability in capturing complex features like mucus plugs and consolidations. To improve the 2D model's performance, a loss adapted to fine structures segmentation was implemented and evaluated, significantly enhancing its accuracy, though not surpassing the 3D model's performance. The models underwent further validation through external evaluation against pulmonary function tests (PFTs), confirming the robustness of the findings. Moreover, this study went beyond comparing metrics; it also included comprehensive assessments of the models' interpretability and reliability, providing valuable insights for their clinical application.

CT evaluation of 2D and 3D holistic deep learning methods for the volumetric segmentation of airway lesions

TL;DR

This study compares 2D and 3D holistic segmentation of cystic fibrosis airway lesions on CT using nnU-Net, introducing a loss tailored to better segment small structures. The 3D model consistently outperforms the 2D counterpart on complex lesions like mucus plugs and consolidations, though a loss-augmented 2D model shows notable improvements without surpassing the 3D results. External validation against pulmonary function tests reveals clinically meaningful correlations for certain lesion volumes, supporting robustness across architectures. Interpretability analyses (Grad-CAM) and Bayesian uncertainty estimation indicate reliable localization and comparable prediction confidence, underscoring the potential of fully automated volumetric scoring for CF severity in clinical practice.

Abstract

This research embarked on a comparative exploration of the holistic segmentation capabilities of Convolutional Neural Networks (CNNs) in both 2D and 3D formats, focusing on cystic fibrosis (CF) lesions. The study utilized data from two CF reference centers, covering five major CF structural changes. Initially, it compared the 2D and 3D models, highlighting the 3D model's superior capability in capturing complex features like mucus plugs and consolidations. To improve the 2D model's performance, a loss adapted to fine structures segmentation was implemented and evaluated, significantly enhancing its accuracy, though not surpassing the 3D model's performance. The models underwent further validation through external evaluation against pulmonary function tests (PFTs), confirming the robustness of the findings. Moreover, this study went beyond comparing metrics; it also included comprehensive assessments of the models' interpretability and reliability, providing valuable insights for their clinical application.
Paper Structure (15 sections, 3 figures, 2 tables)

This paper contains 15 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Overview
  3. Experimental Setup
  4. Data Description
  5. Segmentation model
  6. Loss Function
  7. Internal Evaluation
  8. External Evaluation
  9. Interpretability
  10. Model-agnostic
  11. Data-agnostic
  12. Results and Discussion
  13. Conclusion
  14. Compliance with ethical standards
  15. Acknowledgments

Figures (3)

  • Figure 1: Study Method. CF lesions are color-coded as follows: Red for Bronchiectasis, Green for Peribronchial Thickening, Blue for Bronchial Mucus, Yellow for Bronchiolar Mucus, and Cyan for Consolidation.
  • Figure 2: The comparison of CF lesion segmentations is presented. The first column displays the Ground Truth 3D segmentations, while the second and third columns show the predictions by the modified nnU-Net 3D and 2D models, respectively.
  • Figure 3: The analysis showcases CF lesion feature maps: Ground Truth segmentations in the first row, followed by GradCAM feature maps of modified nnU-Net 3D and 2D models in the second and third rows, respectively. The maps' intensity levels correlate with prediction probabilities, where higher intensity signifies higher probabilities.