C3r-Net: An Edge-Preserving Contrast-Consistency and Calibrated Referral Framework for three-Class Lung Ct Assessment

Authors

  • A. Sirajdeen
  • Dr. S. Rajakumaran

Keywords:

Lung cancer; computed tomography; contrast-consistency learning; gradient-preserving augmentation; ResNet18; confidence calibration; uncertainty referral.

Abstract

Accurate lung computed tomography (CT) analysis requires consistent class decisions and interpretable confidence estimates. Contrast changes can disturb learned representations, while a highly confident output may remain incorrect. This study introduces the Confidence-Calibrated Contrast Consistency Network (C3R-Net), a reproducible methodological framework for investigating both problems. A gradient-conditioned augmentation operator limits tonal perturbations at pronounced image edges. Original and modified views pass through a shared ResNet18 encoder trained using class-weighted cross-entropy and Jensen–Shannon consistency. Post-training temperature scaling supports probability calibration, and an entropy–disagreement rule defines selective referral. Mathematical analysis establishes the operator’s edge-protection property. Original numerical verification using a specified 165-entry label set and normalized probability vectors produces 160 matching assignments (96.97%) for the complete configuration, compared with 155 (93.94%) for the reference configuration. Temperature adjustment changes the calculated expected calibration error from 0.1614 to 0.0400. These computational results verify evaluation arithmetic and decision logic, rather than the predictive accuracy of a fitted CT classifier. An independent clinical validation protocol is therefore specified, including case-disjoint partitioning, matched baseline training, malignant sensitivity, calibration assessment and risk–coverage analysis. C3R-Net contributes an explicitly defined architecture and evaluation pathway for subsequent reproducible lung CT experiments.

Downloads

Published

2026-11-16

How to Cite

Sirajdeen, A., & Rajakumaran, D. S. (2026). C3r-Net: An Edge-Preserving Contrast-Consistency and Calibrated Referral Framework for three-Class Lung Ct Assessment. Adolescência E Saúde, 1132–1139. Retrieved from https://adolescenciaesaude.com/index.php/aes/article/view/3472

Issue

Section

Original Articles