Diagnosis of Pneumoconiosis with Machine Learning

Pneumoconiosis encompasses a group of lung diseases caused by inhaling dust particles. Frequently recognized as an occupational disease, it primarily affects workers in the mining industry. This paper details the use of machine learning algorithms to automate the diagnostic process in two distinct stages: Stage 1 involves lung segmentation, and Stage 2 focuses on classifying X-rays to determine the presence or absence of pneumoconiosis. In Stage 1, a U-Net network is employed for semantic segmentation, achieving an accuracy of 94% on test data and an average accuracy of 98.35% on validation data. Stage 2 introduces a comparative system that complies with the ILO's standard practical guidelines for diagnosis. This stage evaluates four machine learning techniques: Support Vector Machine (SVM), Random Forest, and Naive Bayes and XGBoost. Our findings indicate that dividing the lung into six segments yields the most balanced metrics (including accuracy, precision, F1 score, and recall) across these models. Notably, the XGBoost model outperforms others in this configuration, achieving a remarkable precision of 98%, an accuracy of 90% and a F1 of 84%.