Segmentation-perturbation propagation in pretrained foundation-model embeddings versus IBSI-aligned radiomics for radiologist-assessed lung-nodule malignancy classification on the LIDC-IDRI cohort.

Authors

Abstract

Pretrained foundation-model (FM) embeddings are proposed as robust alternatives to handcrafted radiomic features, yet whether they better withstand segmentation-mask variability, the perturbation that radiomics intraclass-correlation (ICC) filtering exists to address, has not been tested head-to-head. We quantified how segmentation-mask perturbation propagates into feature reproducibility and downstream malignancy discrimination for FM embeddings and Image Biomarker Standardization Initiative (IBSI)-compliant radiomics measured on a common cohort.
Approach. Two 2D FM encoders (BiomedCLIP; RadImageNet ResNet50) were compared against IBSI-compliant radiomics (3D, 107 features; 2D, 102 features) on the full eligible Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) cohort: 556 lung nodules (300 malignant, 256 benign from radiologist consensus ratings, not pathology) from 399 patients. Up to thirteen perturbed masks per nodule (up to four reader masks, eight synthetic, one Dice-calibrated) were generated, and ICC(2,1) was computed across the ten perturbation conditions available for every nodule. Malignancy classification used L2-regularized logistic regression with patient-grouped five-fold cross-validation and bootstrap 95% confidence intervals on every area under the receiver operating characteristic curve (AUC).
Main results. Radiomics outperformed both foundation models on discrimination (AUC 0.940 and 0.930 for radiomics-2D/3D versus 0.833 BiomedCLIP and 0.798 RadImageNet; all radiomics-versus-FM differences significant by DeLong test, p≤10^(-7)). On perturbation stability the foundation models diverged: BiomedCLIP was statistically indistinguishable from radiomics (93.8% versus 90.2-92.5% stable features) and the most stable paradigm, whereas RadImageNet was significantly less stable (75.0%). A pre-specified diameter-stratified analysis showed the radiomics advantage narrows, and partly reverses, within size-matched strata; decision-curve analysis favoured radiomics.
Significance. Two-dimensional foundation-model embeddings offered no discrimination advantage over IBSI radiomics under segmentation perturbation, and greater stability did not translate into greater accuracy. Perturbation stability is therefore decoupled from task-informativeness, and comparative benchmarks should report patient-grouped, size-stratified, calibration-aware metrics, not aggregate accuracy alone.&#xD.

Tags

Topics