Optimized Silicosis Detection Using a Feature-Enriched Multi-Model Ensemble Framework with An Enhanced U-Net Architecture

Silicosis remains a major occupational health concern worldwide. This Research proposes an ensemble learning model combined with an improved U-Net for silicosis detection and classification. The dataset included 1300 healthy lung images and 1328 augmented silicosis images, split 80:20 for training and testing. The ensemble model integrated Inception V2, Inception V3, MobileNet, and ResNet50 for robust classification. Enhanced U-Net with skip connections and optimized hyperparameters achieved a validation accuracy of 98.33% and a Dice coefficient of 0.967. The ensemble outperformed standalone models, achieving an F1-score of 95.57%, a precision of 96.42%, and an MCC of 0.9275. Training was conducted using an NVIDIA i7 CPU, with R-Studio for analytics and visualization. Comparative analysis confirmed the model's superiority, aiding radiologists in faster decision-making.