A Novel and Automated Approach to Classify Radiation Induced Lung Tissue Damage on CT Scans

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Radiation-induced lung damage (RILD) is a frequent complication of thoracic radiotherapy.
• Accurate quantification of RILD is crucial for understanding its progression and impact.

Purpose of the Study:

• To develop and validate an automated framework for segmenting and classifying lung tissue changes post-radiotherapy.
• To create a voxel-wise classification system for lung parenchymal changes, ranging from normal lung to consolidation.

Main Methods:

• A two-stage, active learning-inspired data annotation approach was used for ground truth generation.
• An ensemble of six 2D Unets formed the core of the stage two auto-segmentation algorithm.
• The method was trained on 200 CT scans from 40 patients and tested on 30 CT scans from 6 patients.

Main Results:

• The auto-segmentation achieved high Dice score coefficients (DSC) for normal lung (Class 1: 99%/98%) and consolidation (Class 5: 96%/92%).
• DSC for intermediate tissue classes (2-4) ranged from 26% to 79% on the test set, with lower values attributed to artifacts or rare subgroups.
• Qualitative assessment showed auto-segmentation performance comparable to manual segmentation, with ~90% of cases rated as acceptable by a radiologist.

Conclusions:

• The proposed automated framework for lung tissue class segmentation demonstrates acceptable performance for RILD analysis.
• This tool has the potential to significantly aid large-scale studies investigating RILD patterns and evolution.
• Further refinement may improve accuracy for less frequent or artifact-affected lung tissue changes.