**Title: AReT: Anatomy-Regularized TensoRF for Sparse-View Lung Nodule Volumetry Using Digitally Reconstructed Radiographs

Abstract:

This study addresses a critical, previously undocumented limitation in applying TensoRF to X-ray attenuation fields. We found that the default density shift parameter of -10—originally designed for RGB scene reconstruction—effectively dampens density gradients. This suppression renders sparse-view medical reconstruction impossible, irrespective of the learning rate or regularization techniques employed. By resetting the density shift to zero, we restored gradient flow, thereby enabling stable volumetric reconstruction of pulmonary nodules from just three orthogonal X-ray projections.

Leveraging this finding, we introduce AReT, an anatomy-regularized tensorial radiance field framework tailored for lung nodule reconstruction. Utilizing coronal, sagittal, and axial projections from the LIDC-IDRI dataset (which includes radiologist-annotated nodules in 19 patients), AReT is specifically engineered for sparse-view thoracic imaging. It integrates chest-anatomy-aware regularization that merges L1 sparsity with total variation smoothness. This stands in contrast to existing NeRF methods, which typically demand dense multi-view acquisition.

Our systematic evaluation across 11 reconstruction strategies demonstrates that anatomy-aware regularization consistently surpasses approaches guided by generative priors. When benchmarked against radiologist consensus segmentations, AReT delivered a Pearson correlation coefficient of r=0.983 for nodules with a volume of 10 mm (n=14). The model exhibited a median absolute volumetric error of 11.4% and near-zero systematic bias of -77.3 mm^3, representing an 8.4-fold improvement over spherical volume approximation.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC