Evaluating the robustness of features generated by a foundation model from CT with different reconstruction...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiomics

Background:

• Foundation models offer generalized imaging features for diagnostic tasks.
• Clinical use of imaging biomarkers is limited by acquisition/reconstruction variability.
• Evaluating foundation model robustness in low-dose computed tomography (LDCT) is crucial.

Purpose of the Study:

• Assess the robustness of foundation model-extracted features from LDCT scans under diverse image conditions.
• Compare the performance of foundation model features against Pyradiomics for nodule outcome classification.
• Investigate the impact of image variations on feature stability and predictive power.

Main Methods:

• Extracted features using a pretrained foundation model from 59 LDCT scans with 6 image variations (reconstruction kernels, slice thicknesses).
• Evaluated feature robustness using concordance correlation coefficient (CCC) against a reference condition.
• Assessed nodule outcome classification performance and compared with Pyradiomics.

Main Results:

• Foundation model features demonstrated high robustness with mean CCC values ranging from 0.937 to 0.984 across conditions.
• Foundation model features outperformed Pyradiomics in nodule outcome classification.
• Performance inconsistencies across image conditions were observed.

Conclusions:

• Foundation models provide robust imaging features for LDCT, outperforming traditional methods.
• Further research into harmonization and modeling is needed to maximize clinical utility.
• Enhanced generalizability and consistency are key for widespread adoption of imaging biomarkers.