ADAPTIVE CT PREPROCESSING AND ROI STANDARDIZATION FOR PULMONARY NODULE REPRESENTATION

Abstract

Preprocessing is part of almost every thoracic computed tomography (CT) pipeline for pulmonary nodule analysis, yet its individual steps are rarely measured on their own. This paper evaluates seven preprocessing stages on 30 CT scans from LIDC-IDRI containing 86 radiologist-annotated nodules (37 benign, 26 indeterminate, 23 malignant): Hounsfield unit conversion, isotropic resampling, lung segmentation, intensity standardization, denoising, contrast enhancement, and region-of-interest (ROI) extraction. Two adaptive components are tested, a lung-restricted, noise-adaptive form of contrast-limited adaptive histogram equalization (CLAHE) and a size-adaptive ROI. Selected parameters are chosen by patient-level five-fold cross-validation, and uncertainty is reported as patient-level bootstrap intervals. A fixed lung window kept scans more consistent than per-scan min–max scaling (pairwise histogram distance 0.085 against 0.235). A corrected threshold-based lung mask retained 96.5% of nodules at a mean lung volume of 4.8 L, whereas the mask of the original pipeline covered 86% of the scan. A cross-validated bilateral filter raised structural similarity in the lungs from 0.910 to 0.955 at 25 HU of added noise and raised nodule contrast-to-noise ratio (CNR) by 7.6% without loss of edge sharpness, although at 10 HU no filtering scored higher. CLAHE lowered CNR at every clip limit above 1 (median ratio 0.79 at clip limit 2); the adaptive variant was selected at its weakest setting and did not improve on applying no enhancement (median CNR ratio 0.998, mean 0.919). The size-adaptive ROI reduced the coefficient of variation of nodule occupancy from 2.06 to 0.63 with no truncation. The study shows which steps contribute to a standardized nodule representation and which do not.