A 3D AI system for automated detection of pulmonary infected regions from HRCT scans

Abstract

Interstitial lung diseases (ILDs) are a diverse group of pulmonary disorders that cause progressive scarring and inflammation of the lung tissue. Accurate diagnosis is challenging due to the heterogeneous radiological appearance of lesions and the high volume of HRCT data requiring expert interpretation. This complexity underscores the need for automated systems capable of reliable and efficient ILD analysis. This Dissertation presents a modular 3D deep learning system for the automated detection and classification of ILD-related lesions from high-resolution computed tomography (HRCT) scans. The pipeline is composed of three main stages: initial lung segmentation using a 3D U-Net model, binary lesion detection using a patch-based 3D CNN classifier (Simple3DCNN), and a second classification stage that performs multi-class lesion categorization across the most common lesions using a fine-tuned version of the same CNN architecture. To improve interpretability and trustworthiness of the system's predictions, 3D Grad-CAM (Gradient- weighted Class Activation Mapping) was applied to highlight salient regions influencing the model's classification decisions. The system was trained and validated on a carefully preprocessed dataset, with patch sampling strategies, and class balancing techniques. The lung segmentation model achieved excellent results (Dice coefficient: 0.99, Hausdorff distance: 3.17), while the binary lesion detector reached high sensitivity (0.993) and accuracy (0.994). The multi-class classification stage achieved an overall accuracy of 88.73% and macro F1-score of 88.7% across most com- mon lesion types. To validate spatial reasoning, Grad-CAM heatmaps were overlaid on the original HRCT patches, confirming the network's attention to clinically relevant regions and structures. These results demonstrate the pipeline's ability to accurately detect and differentiate ILD lesions in 3D space, while also providing visual interpretability through Grad-CAM that can enhance clinical confidence in automated decision-making. The system provides a solid foundation for future integration into diagnostic workflows and further extension toward real-time, explainable AI tools in pulmonary imaging. Key words: Interstitial Lung Diseases (ILD), 3D Medical Image Segmentation, Pulmonary Lesion Classification, Deep Learning in Thoracic Imaging