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

Md. Enamul Hoq, Sharafat Hossain, Imraul Emmaka, Linda Larson-Prior, Lawrence Tarbox, Jonathan Bona, Donald Johann Jr., Fred Prior

Keywords: lung CT, NLST, 2.5D, CNN, Vision Transformer, failure modes

TL;DR: 2.5D CNNs offer the best trade-off between performance, stability, and efficiency for lung CT classification, while full 3D and transformer models often introduce instability without consistent gains.

Registration Requirement: Yes

Abstract:

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 comparing architectures, we study how input dimensionality (2D, 2.5D, 3D) affects model behavior across convolutional neural networks (CNNs) and Vision Transformers (ViTs).
Using a leakage-free NLST cohort ($n=1{,}977$) with supporting LIDC-IDRI data, we find that 2.5D CNN achieves the best discrimination (ROC-AUC 0.682, 95\% CI [0.546, 0.799]) with stable operating behavior. In contrast, 3D CNNs show threshold instability, and transformers exhibit degenerate predictions (e.g., all-positive).
Our results demonstrate that dimensionality governs both performance and failure mode. For lung cancer screening classification, 2D and 2.5D provide a more reliable trade-off between performance, stability, and computational efficiency than full 3D representations.

Reproducibility: The codebase for this work is publicly available at: https://github.com/Emmaka9/when-is-3d-worth-it The repository includes scripts for data preprocessing, model training (2D, 2.5D, and 3D CNNs and ViTs), evaluation, and generation of all reported metrics. We provide detailed instructions to reproduce the experimental pipeline, including dataset preparation, training configurations, and evaluation protocols. All experiments are designed to be reproducible under a leakage-free patient-level split using the NLST dataset, with supporting experiments on LIDC-IDRI.

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 45