Calibrated ensemble deep learning for four-class lung CT image classification and uncertainty-based referral

Background Deep-learning systems may support chest CT interpretation, but their reliability depends on balanced class performance, calibrated probabilities, and transparent handling of uncertain predictions. We developed an ensemble framework for four-class image-level classification of adenocarcinoma, large cell carcinoma, squamous cell carcinoma, and normal lung images. Methods We retrospectively analyzed 1,000 publicly available chest CT images using the dataset-provided training, validation, and test partitions. ImageNet-pretrained Xception, DenseNet121, and ResNet50V2 models were trained with augmentation, random oversampling, and focal loss. Validation-selected models were combined in a weighted ensemble, temperature calibrated, decision scaled, and evaluated for discrimination, calibration, predictive entropy, and uncertainty-based referral. Results On the test partition, the calibrated decision-scaled ensemble achieved 77.8% accuracy (95% CI, 73.7%–82.7%), 83.5% balanced accuracy, a macro F1 of 0.812, and a macro AUROC of 0.980. Expected calibration error decreased from 0.084 to 0.062. Predictive entropy was higher for errors (p<0.001), and deferring the most uncertain 10% increased accuracy to 83.0%. Conclusion The calibrated ensemble showed strong probability-based discrimination but uneven subtype classification, particularly between adenocarcinoma and squamous cell carcinoma. Predictive entropy identified error-prone images and enabled improved performance after selective referral. These findings support uncertainty-aware assistance while underscoring the need for patient-level, multicenter external validation before clinical implementation.