Machine Learning Expert System for Tuberculosis Analysis

Abstract

Pulmonary tuberculosis still represents a significant public health challenge, especially in regions with poor medical infrastructure. Considering the advancement of Artificial Intelligence techniques and the need for accessible solutions, this work aims to develop an embedded expert system to assist in the automated detection of tuberculosis through radiographic images. The methodology adopted utilized the Edge Impulse platform and the MobileNetV2 model with transfer learning, applied to a public database containing 4,199 chest radiographs labeled as either "with tuberculosis" or "without tuberculosis". The model was trained with data augmentation and quantization optimization techniques, aiming for greater computational efficiency in low-cost devices. Tests demonstrated promising results, achieving an accuracy of 98.48%, an area under the receiver's operational characteristic curve of 0.98, and an F1-score of 0.99, with superior performance even in environments with processing constraints. The comparison between the quantized and non-quantized models confirmed the viability of embedded use without significant performance loss.