A Data-Driven Multiscale Nonlinear Segmentation Framework for Silicosis Detection Using Machine Learning

Early-stage silicosis is challenging to identify since lung alterations might appear as micronodules, distorted parenchymal patterns, or fibrotic changes. In this paper, we introduce a nonlinear multiscale image segmentation framework, with machine-learning classification, for computer-aided assessment of silicosis. The proposed method combines nonlinear total-variation (TV)-flow scale-space construction, hierarchical region merging, data-driven rgion pruning, and region-level machine learning classification. In contrast to traditional image-based classifiers, the approach generates interpretable lung subregions featuring micronodular and fibrotic lesions prior to classification. A nonlinear scale-space stack is constructed by using diffusion-based partial differential equations while maintaining lesion boundaries at different anatomical scales. The resulting structures are connected via a hierarchical region tree, and a data - driven pruning procedure is employed to prune unstable or non-anatomical regions within each path without any manual intervention. Intensity, shape, texture, and multiscale features are extracted from the retained regions and then used for supervised classification. The study consisted of 64 subjects and 286 HRCT slices. The segmentation-based method performed 9–11 % better in classification accuracy compared to the single-scale approaches or non-oriented features and achieved ~ 92 % sensitivity in detecting early stage silicosis. The framework also showed better edge preservation, less over segmentation, and practical computational efficiency. These results motivate the continued investigation of the approach as an interpretable computer-aided silicosis assessment framework.