Virtual-Eyes: Quantitative Validation of a Lung CT Quality-Control Pipeline for Foundation-Model Cancer Risk Prediction

TL;DR

The study addresses the lack of quantified preprocessing in LDCT cancer-risk pipelines by introducing Virtual-Eyes, a deterministic 16-bit, lung-focused QC that enforces $512\times512$ geometry and contiguous lung blocks. By evaluating four architectures—two generalist FMs (RAD-DINO, Merlin) and two specialists (Sybil, ResNet-18)—on $765$ NLST patients, the authors show that anatomically standardizing inputs improves a generalist encoder (notably increasing slice- and patient-level AUC and calibration for RAD-DINO) but can impair specialist models that rely on broader contextual cues or domain-specific pretraining (Sybil, ResNet-18, Merlin). The findings argue for model-specific preprocessing choices, suggesting that QC can stabilize and improve generalist workflows while exposing shortcuts in specialized models and domain-mismatched foundation models. Overall, Virtual-Eyes offers a low-cost, reproducible front end to harmonize LDCT inputs for foundation-model pipelines and highlights the ongoing need for domain-aware pretraining and fine-tuning in medical imaging.

Abstract

Robust preprocessing is rarely quantified in deep-learning pipelines for low-dose CT (LDCT) lung cancer screening. We develop and validate Virtual-Eyes, a clinically motivated 16-bit CT quality-control pipeline, and measure its differential impact on generalist foundation models versus specialist models. Virtual-Eyes enforces strict 512x512 in-plane resolution, rejects short or non-diagnostic series, and extracts a contiguous lung block using Hounsfield-unit filtering and bilateral lung-coverage scoring while preserving the native 16-bit grid. Using 765 NLST patients (182 cancer, 583 non-cancer), we compute slice-level embeddings from RAD-DINO and Merlin with frozen encoders and train leakage-free patient-level MLP heads; we also evaluate Sybil and a 2D ResNet-18 baseline under Raw versus Virtual-Eyes inputs without backbone retraining. Virtual-Eyes improves RAD-DINO slice-level AUC from 0.576 to 0.610 and patient-level AUC from 0.646 to 0.683 (mean pooling) and from 0.619 to 0.735 (max pooling), with improved calibration (Brier score 0.188 to 0.112). In contrast, Sybil and ResNet-18 degrade under Virtual-Eyes (Sybil AUC 0.886 to 0.837; ResNet-18 AUC 0.571 to 0.596) with evidence of context dependence and shortcut learning, and Merlin shows limited transferability (AUC approximately 0.507 to 0.567) regardless of preprocessing. These results demonstrate that anatomically targeted QC can stabilize and improve generalist foundation-model workflows but may disrupt specialist models adapted to raw clinical context.

Figures (11)

• Figure 1: Overall Virtual-Eyes workflow. (a) Lung-focused preprocessing that rejects non-lung slices and extracts a contiguous 16-bit LDCT lung block. (b) Foundation and specialist models (RAD-DINO, Merlin, Sybil, ResNet-18) used to compute slice-/series-level embeddings, which are pooled into patient-level representations and fed to an MLP classifier for lung cancer risk prediction.
• Figure 2: RAD-DINO on the test set. (a) Patient-level ROC (mean pooling) for Raw vs. Virtual-Eyes. (b) Slice-level ROC. (c) Bland--Altman plot of patient-level mean probabilities (Virtual-Eyes minus Raw), showing a small positive bias and tighter agreement.
• Figure A1: Representative Virtual-Eyes preprocessing example showing an accepted lung block. Non-lung slices (neck and abdomen) are rejected, and the retained block preserves the full 16-bit LDCT grid.
• Figure A2: Additional Virtual-Eyes visualization across multiple series, highlighting common rejection patterns such as very short scout views, incomplete chest coverage, and incorrect field-of-view.
• Figure B1: RAD-DINO embedding structure for Raw vs. Virtual-Eyes inputs. (a) t-SNE visualization of preprocessed vs. raw embeddings. (b) UMAP visualization of preprocessed vs. raw embeddings. Virtual-Eyes yields tighter, better separated clusters of cancer and non-cancer slices.