When is 3D Worth It? A Resource-Performance Frontier for CNNs and Transformers in Lung CT

Three-dimensional models are widely assumed preferable for volumetric medical imaging, yet their practical value depends on whether performance gains justify added computational cost and complexity. Rather than proposing a new architecture, we study how input dimensionality (2D, 2.5D, 3D) affects model behavior across convolutional neural networks (CNNs) and Vision Transformers (ViTs) under a fixed training protocol. Using a leakage-free NLST cohort (n = 1,977) with supporting LIDC-IDRI data, we find that the 2.5D CNN offers the most favorable discrimination-stability trade-off in our comparison (ROC-AUC 0.682, 95% CI [0.546, 0.799]) with a stable operating point. In contrast, 3D CNNs show threshold instability, and transformers exhibit degenerate predictions, such as all-positive predictions. Confidence intervals are wide and overlapping, so we present these results as a controlled resource-performance frontier and a failure-mode taxonomy rather than as definitive superiority claims. For class-imbalanced lung cancer screening classification, 2D and 2.5D inputs provide a more reliable trade-off between performance, stability, and computational efficiency than full 3D representations.

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

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.