Source-linked AI summary
Progressive and Multi-Path Holistically Nested Neural Networks for Pathological Lung Segmentation from CT Images
Adam P. Harrison, Ziyue Xu, Kevin George, Le Lu, Ronald M. Summers, Daniel J. Mollura
TL;DR
Pathological lung segmentation must accommodate variation in disease appearance and shape, and shape- or localization-based approaches can require accounting for those variations. The paper proposes P-HNNs, a bottom-up FCN with progressive multi-path connections, and reports strong performance across 929 CT scans.
Problem
PLS must handle wide pathological appearance variability, while shape- or landmark-based approaches require accounting for anatomical variation.
Source: §1 · p. 2Method
P-HNNs adapt HNNs into a bottom-up FCN with progressive multi-path connections that combine stage activations without extra parameters over the base FCN.
Source: §1 · p. 2, §2 · pp. 4–5Results
Takeaways & Limitations
P-HNNs offer a simple, effective means of producing robust pathological lung segmentation masks.
Source: §4 · p. 7A data-archiving issue limited comparison with Mansoor et al. to 88 available CT scans, only 47 of which had PLS masks from their method.
Source: §3 · p. 5
Abstract
from arXiv · showhide
Pathological lung segmentation (PLS) is an important, yet challenging, medical image application due to the wide variability of pathological lung appearance and shape. Because PLS is often a pre-requisite for other imaging analytics, methodological simplicity and generality are key factors in usability. Along those lines, we present a bottom-up deep-learning based approach that is expressive enough to handle variations in appearance, while remaining unaffected by any variations in shape. We incorporate the deeply supervised learning framework, but enhance it with a simple, yet effective, progressive multi-path scheme, which more reliably merges outputs from different network stages. The result is a deep model able to produce finer detailed masks, which we call progressive holistically-nested networks (P-HNNs). Using extensive cross-validation, our method is tested on multi-institutional datasets comprising 929 CT scans (848 publicly available), of pathological lungs, reporting mean dice scores of 0.985 and demonstrating significant qualitative and quantitative improvements over state-of-the art approaches.
1 Introduction
Pathological lung segmentation must handle wide variation in abnormal appearance, while shape-dependent approaches may struggle when anatomy varies. The introduction presents P-HNNs as a simple bottom-up approach and evaluates it on a large multi-institutional CT collection.
1 Introduction
- PLS must handle appearance variability caused by abnormalities, while reliance on 3D shape or landmark priors can leave methods vulnerable to unaccounted shape variation.The introduction gives a one-lung case as an example of a challenging shape variation for top-down methods.Source: §1 · p. 2
- Existing fully bottom-up PLS methods show promise but can struggle with severe pathologies, motivating investigation of deep fully convolutional networks for segmentation.The cited FCN work is described as state of the art in computer vision and medical imaging segmentation [7].Source: §1 · p. 2
- The paper proposes P-HNNs, a bottom-up FCN approach that adapts HNNs with progressive multi-path constraints to merge outputs simply and robustly.Unlike complex coarse-to-fine multi-path methods, the proposed constraint requires no additional convolutional layers or network parameters.Source: §1 · p. 2
- 0.985±0.011 average Dice was achieved in five-fold cross-validation on 929 challenging thoracic CT scans spanning infection-, ILD-, and COPD-based pathologies.846 scans were publicly available for future comparisons.Source: §1 · p. 2
2 Methods
P-HNNs extend HNN deep supervision with progressive multi-path connections that refine side outputs and eliminate the inconsistent fused output. Their implementation uses 2D axial CT slices.
- HNNs compute side-output losses at intermediate stages and merge predictions from different stages, atop a CNN divided into pooling-separated stages.Source: §2 · p. 3
- A global class-balancing weight β uses a training-set sample balance and is reported to outperform Xie et al.’s image-specific scheme.Source: §2 · p. 4
- P-HNNs combine current- and previous-stage activations before sigmoid, so each side output can add to or subtract from its predecessor.Source: §2 · pp. 4–5
- The simpler progressive links contrast with coarse-to-fine designs that add parameters near the original FCN path; P-HNNs use fewer parameters than standard HNNs.Source: §2 · pp. 4–5
- This progressive refinement removes HNN’s fused output, which can be inconsistent and sometimes worse than individual side outputs.Source: §2 · pp. 4–5
- The implementation uses 2D axial slices with three HU windows; the authors choose 2D to retain large field of view and accommodate variable slice thickness.Source: §2 · p. 5
3 Experiments
Five-fold cross-validation showed that P-HNN outperformed standard HNN on the full test set and Mansoor et al.’s method on 47 NIH cases. The evaluation used multi-institutional CT data and included qualitative examples of segmentation differences.
3 Experiments
- The five-fold evaluation used 773 LTRC, 73 UHG, and 83 NIH scans, with folds separated by patient and dataset.Training used every tenth LTRC slice and all slices from the other datasets; scan-mask pairs with annotation errors were excluded.Source: §3 · p. 5
- Only 47 of the 88 shared NIH scans had masks produced by Mansoor et al., constraining that comparison to those cases.The archive contained 88 CT scans, but only 47 corresponding PLS masks from the comparator method.Source: §3 · p. 5
- Examples show P-HNN capturing bullae, infiltrates, and finer lung details, while both methods included the right main bronchus in one failure case.Against Mansoor et al., P-HNN examples also show better lung-field coverage and ground-glass opacity segmentation, while the comparator leaked into the esophagus and intestine.Source: Figure 2 · p. 6, Figure 3 · p. 7
- On the entire test dataset, P-HNN achieved DS and ASD values of 0.985 and 0.762 mm, outperforming standard HNN's 0.978 and 1.063 mm, respectively (p < 0.001, Wilcoxon signed-rank test).Source: §3 · p. 6, Table 1 · p. 6
- P-HNN significantly outperformed Mansoor et al.’s method in Dice score on 47 NIH volumes, while its better average surface distance was not statistically significant.The comparison used a subset with infectious diseases; Fig. 4(b) shows the score-distribution comparison.Source: §3 · p. 6, Table 1 · p. 6, Figure 4 · p. 7
4 Conclusion
P-HNN combines deep supervision and multi-path connections to address FCN resolution coarsening without adding parameters over the base FCN. Across 929 thoracic CT scans, it outperformed standard HNN and improved masks against a state-of-the-art tool.
4 Conclusion
- P-HNN consistently outperformed standard HNN (p < 0.001), with mean Dice score 0.985 ± 0.011 across 929 thoracic CT scans.The scans covered infection-, ILD-, and COPD-based pathologies.Source: §4 · p. 7
- Progressive multi-path enhancement addresses FCN coarsening resolution without adding parameters over the base FCN.P-HNN combines deep supervision with multi-path connections.Source: §4 · p. 7
- P-HNN produced significantly improved pathological lung masks compared with a state-of-the-art tool.The conclusion identifies Mansoor et al.’s method as the comparison tool.Source: §4 · p. 7
- The authors state that P-HNN could also be applied to other pathological lung morbidities and segmentation tasks.Source: §4 · p. 7