Uncertainty-aware lung CT slice classification using a lightweight CNN with selective referral

Abstract Lung CT slice classification may support image-level decision assistance, but deterministic neural networks do not explicitly indicate when a prediction is unreliable. This study presents a lightweight uncertainty-aware convolutional neural network (CNN) for three-class classification of benign, malignant, and normal CT slices, integrating Monte Carlo (MC) Dropout, validation-selected selective referral, and Grad-CAM visualization. After removing 43 exact duplicate files, 1,054 IQ-OTH/NCCD images were analyzed using an image-level split of 842 training, 106 validation, and 106 test images (seed 42). Because reliable patient identifiers could not be recovered, the study is considered an internal image-level proof-of-concept rather than patient-independent diagnostic validation. With 50-pass MC Dropout, the model correctly classified 105/106 test slices, achieving 99.06% accuracy (exact 95% CI: 94.86–99.98%) and a weighted F1-score of 99.04%. The single misclassified benign slice was predicted as normal and exhibited high predictive uncertainty. Selective referral was based on the scalar uncertainty score $${\text{U}} = {\text{max}}\left( {1 - {\text{Confidence}},\,{\text{H}}_{{{\text{norm}}}} ,\,{\text{MI}}_{{{\text{norm}}}} ,\,{\text{VR}}} \right)$$ , with a threshold selected as the 80th percentile of validation-set U values $$\left( {\tau_{U} = 0.174166} \right)$$ , without using test labels or test performance. This operating point accepted 88/106 slices (83.02% coverage) and referred 18/106 (16.98%), with an observed accepted accuracy of 100% (95% CI: 95.89–100%) and an observed selective risk of 0% (95% CI: 0–4.11%). Calibration analysis yielded a 10-bin equal-width expected calibration error of 0.0295, a multiclass Brier score of 0.0137, and a negative log-likelihood of 0.0385. Deep Ensemble and Ensemble MC Dropout each correctly classified all 106 test slices, although these results remain subject to the same image-level data limitation. Simulated low-contrast and Gaussian-noise perturbations reduced accuracy and increased referral, indicating responsiveness of the uncertainty score to distributional degradation. A separately retrained binary lightweight CNN variant on the Chest CT-Scan Images Dataset correctly classified all 286 Cancer-versus-Normal test slices (100%; 95% CI: 98.72–100%), but its validation-selected referral rule accepted only 134/286 slices (46.85%) and referred all 25 Normal test slices. Therefore, the second-dataset experiment should be interpreted as a dataset-specific image-level benchmark rather than direct external validation of the primary three-class model. These findings support further patient-wise, multi-center evaluation of uncertainty-aware CT-slice classification and selective referral.