AReT: New AI model reconstructs lung nodules from just 3 X-ray images

A new paper from researchers Spoorthi M and Suja Palaniswamy introduces AReT (Anatomy-Regularized TensoRF), a radiance field framework designed for sparse-view lung nodule volumetry. The authors first identify a critical failure mode in existing TensoRF models: the default density shift of -10, originally intended for RGB scene reconstruction, suppresses density gradients and prevents stable volumetric reconstruction from sparse X-ray projections. By setting the density shift to zero, they restore gradient flow and enable reconstruction from only three orthogonal X-ray views—coronal, sagittal, and axial—instead of the dozens typically required by NeRF-based methods. The model is trained on the LIDC-IDRI dataset (19 patients with radiologist-annotated nodules) and incorporates chest-anatomy-aware regularization combining L1 sparsity and total variation smoothness.

Systematic comparisons against 11 reconstruction strategies show that AReT's anatomy-aware regularization outperforms generative-prior-guided approaches. Evaluated against radiologist consensus segmentations, AReT achieves a Pearson correlation coefficient of 0.983 (p<0.0001) for clinically actionable nodules ≥10 mm (n=14), with a median absolute volumetric error of just 11.4%, near-zero systematic bias of -77.3 mm³, and an 8.4× improvement over spherical volume approximation. This advance could enable accurate lung nodule monitoring with minimal X-ray exposure, benefiting early cancer detection and reducing patient radiation dose.