High-accuracy lung region segmentation and injury classification using deep networks

This study developed and evaluated a hybrid deep learning framework for automated lung segmentation and multi-class severity classification in chest CT scans, addressing key methodological gaps including heterogeneous dataset labeling, pseudo-label bias in segmentation, limited interpretability of intensity thresholds, and unvalidated preprocessing effects. A dataset of 3,921 chest CT images from three geographically distinct sources (Iran, Bangladesh, Iraq) was harmonized into four severity classes (No Lesion, Mild, Moderate, Severe) through a structured expert annotation protocol (κ = 0.847) and augmented to 4,392 balanced samples. U-Net architectures with three CNN backbones (VGG16, ResNet50, Xception) were trained for lung segmentation with CLAHE preprocessing and evaluated against both morphological pseudo-labels and independently annotated expert ground truth. A novel Brightness Intensity Range (BIR) classifier analyzed pixels within the normalized intensity range [0.35, 0.70], corresponding to approximately − 450 to + 100 HU and encompassing ground-glass opacities and partial consolidation, using perimeter-to-area ratio thresholds for severity stratification. Hybrid classification pipelines combined CNN feature extraction with SVM, Gradient Boosting, and Logistic Regression classifiers. A controlled ablation study isolated the contribution of each preprocessing step. All models were evaluated using patient-level fivefold cross-validation to prevent data leakage between training and test partitions. The VGG16-based U-Net achieved 98.1% segmentation accuracy with a Dice coefficient of 0.927 against pseudo-labels and 0.921 ± 0.089 against expert annotations, confirming that pseudo-label training introduced no systematic bias. ResNet50 + GradientBoosting achieved the highest classification accuracy of 96.35 ± 0.82% with ROC-AUC exceeding 99% across all classes. The BIR-VGG16 model reached 94.8% accuracy with ROC-AUC values of 0.963 to 0.988. The ablation study identified normalization as the dominant preprocessing contributor (34.3% accuracy gain), followed by CLAHE (21.7%) and resizing (6.0%), mechanistically explaining the large performance gap between preprocessed and non-preprocessed conditions. Cross-dataset generalization yielded 93.78% accuracy on an independent cohort. The proposed framework demonstrates superior performance, methodological transparency, and clinical viability. The explicit HU-grounded BIR thresholds, expert-validated segmentation, and ablation-confirmed preprocessing rationale collectively enhance the generalizability of the approach for automated lung pathology assessment.