Edit this page

NA-MIC Project Weeks

Back to Projects List

3D Slicer Lung CT Segmentation and Analysis

Key Investigators

Presenter location: Online

Project Description

This is a follow-up to previous 3D Slicer lung CT segmentation PW projects.

Objective

To improve the LungCTAnalysis extension analysis in 3D Slicer,

which is frequently used (40 runs per day)

image

image

the following steps could be taken:

  1. Improve vessel segmentation

  2. Develop a better concept for lung segment (sublobar) segmentation in 3D Slicer

  3. Identify tumors belonging to segments and consider safety margins

  4. Suggest resection of segments which include nutritive vessel resection for neighboring tumors

  5. Differentiate pulmonary arteries and veins reliably

  6. Work on current OpenSourceCOVID publication

Approach and Plan

Work on a dedicated Angio-CT.

Progress and Next Steps

We analyzed the usage statistics of Lung CT Analyzer and decided to keep the mechanism.

Lung CT Segmenter

The most expedient way to obtain a precise surgical planning result at the moment involves the use of combined airway segmentation and vessel volume rendering, utilizing the centerblock technique. This method requires minimal manual intervention. We have chosen to enhance the division of vascular structures within the vessel mask by eliminating the hilar structures, where there’s a prominent overlay of vessels. Subsequently, we employ a ‘grow-from-seeds’ analysis which promises improved accuracy. The VMTK centerline analysis, however, appears to be a less preferable option. This is largely due to its lack of efficiency in differentiating between pulmonary arteries and vessels.

Lung CT Analyzer

The primary constraint of the computational method appears to be its sensitivity to the selected threshold values. Even minor variances in well-calibrated clinical CTs can lead to appreciable differences in the final outcome, affecting about 20-30% of cases. The likely source of this issue is the use of rigid threshold values, which inherently makes the classification highly responsive to the specific threshold set. Andras Lasso has proposed mitigating this sensitivity through the application of soft classifiers. We intend to follow this approach, potentially augmenting it with AI pattern detection.

Illustrations

image

Fig 1: Pulmonary vesselmask

image

Fig 2: Combined airway segmentation and vessel volume rendering, centerblock technique

image

Fig 3: Centerline analysis with VMTK

Background and References

Lung CT Analyzer

https://github.com/rbumm/SlicerLungCTAnalyzer

From Voxels to Prognosis: AI-Driven Quantitative Chest CT Analysis Forecasts ICU Requirements in 78 COVID-19 Cases (preprint)

https://doi.org/10.21203/rs.3.rs-3027617/v3