Harnessing Advanced Machine Learning Techniques for Microscopic Vessel Segmentation in Pulmonary Fibrosis Using Novel Hierarchical Phase-Contrast Tomography Images

Abstract

Background: Fibrotic lung disease is a progressive illness that causes scarring and ultimately respiratory failure, with irreversible damage by the time it is diagnosed on computed tomography imaging. Recent research postulates the role of the lung vasculature on the pathogenesis of the disease. With the recent development of high-resolution hierarchical phase-contrast tomography (HiP-CT), we have the potential to understand and detect changes in the lungs long before conventional imaging. However, to gain quantitative insight into vascular changes you first need to be able to segment the vessels before further downstream analysis can be conducted. Aside from this, HiP-CT generates large-volume, high-resolution data which is time-consuming and expensive to label.

Objectives: This project aims to qualitatively assess the latest machine learning methods for vessel segmentation in HiP-CT data to enable label propagation as the first step for imaging biomarker discovery, with the goal to identify early-stage interstitial lung disease amenable to treatment, before fibrosis begins.

Methods: Semisupervised learning (SSL) has become a growing method to tackle sparsely labeled datasets due to its leveraging of unlabeled data. In this study, we will compare two SSL methods; Seg PL, based on pseudo-labeling, and MisMatch, using consistency regularization against state-of-the-art supervised learning method, nnU-Net, on vessel segmentation in sparsely labeled lung HiP-CT data.

Results: On initial experimentation, both MisMatch and SegPL showed promising performance on qualitative review. In comparison with supervised learning, both MisMatch and SegPL showed better out-of-distribution performance within the same sample (different vessel morphology and texture vessels), though supervised learning provided more consistent segmentations for well-represented labels in the limited annotations.

Conclusion: Further quantitative research is required to better assess the generalizability of these findings, though they show promising first steps toward leveraging this novel data to tackle fibrotic lung disease.

Fig. 1

Example of 2.5-μm resolution Synchrotron data of the lung. (A) Biopsy sample with PPFE. (B) Biopsy sample with PPFE and IPF. (C) Sparsely annotated sample with PPFE and IPF. IPF, idiopathic pulmonary fibrosis; PPFE, pleuroparenchymal fibroelastosis.

Fig. 2

Workflow of the MisMatch framework for semisupervised segmentation. The encoder processes the input image, generating features that are passed to two parallel decoders: the positive attention shifting decoder ( ) for dilated feature prediction and the negative attention shifting decoder ( ) for eroded feature prediction. For labeled data, supervised loss (Dice Loss) is calculated, while for unlabeled data, consistency regularization (MSE Loss) is applied between the outputs of and . The final segmentation prediction is obtained by averaging the outputs of the two decoders.

Fig. 3

Workflow of the SegPL algorithm with a fixed threshold for pseudo-label generation. Input data are split into labeled and unlabeled datasets. In the E-step, pseudo-labels ( ) are generated for unlabeled data using the model's predictions ( θ ) with a fixed threshold ( T  = 0.5). In the M-step, the model is refined by optimizing a combined loss function ( L _total ) comprising supervised Dice Loss ( L _L ) for labeled data and unsupervised Dice Loss ( L _U ) for pseudo-labeled data. The process iterates until convergence, yielding the final segmentation output.

Fig. 4

MisMatch segmentation overlay (red) on incomplete ground truth labels (green) from dataset B using Design method 1. Of note, a vessel in the top left-hand corner (black arrow) was not labeled in the ground truth, as well as a vessel in the middle (blue arrow). There is incomplete labeling of the ground truth. There is partial labeling of the two noted vessels not in the ground truth, with minimal additional false positive labels.

Fig. 5

MisMatch segmentation overlay (blue) on incomplete ground truth labels (green) from dataset B using Design Method 2. Incomplete labeling of larger ground truth vessel. More complete labeling of the two noted vessels not in the ground truth compared with the other training strategy, as well as other small vessels, however further additional false positive labels.

Fig. 6

SegPL segmentation overlay (yellow) on incomplete ground truth labels (green) from dataset B using Design Method 1. There is a labeling of the two noted vessels not in the ground truth, as well as other small vessels, but with some additional false positive labels, more than seen in MisMatch.

Fig. 7

SegPL segmentation overlay (pink) on incomplete ground truth labels (green) from dataset B using Design Method 2. Prediction from SegPL with validation (pink). There is labeling of the 2 vessels not in the ground truth, as well as other small vessels, but with some additional false positive labels.

Fig. 8

nnU-Net segmentation overlay (blue) on ground truth labels (green) from dataset B. Good segmentation of the ground truth label and of the small vessel in the top left-hand corner which was not originally labeled. However, smaller vessels are not labeled.