Mobile App for Tuberculosis Detection Using Deep Learning and NLP-Based Recommendations

Tuberculosis remains one of the leading causes of mortality from infectious diseases and conventional diagnostic methods are often limited by subjectivity. This study presents a mobile application for automated tuberculosis diagnosis, integrating a convolutional neural network (CNN) for chest X-ray classification with a natural language processing (NLP) module for clinical recommendations. Three CNN architectures were evaluated (DenseNet121, ResNet50, and MobileNetV2) using the CRISP-DM methodology, with a Kaggle dataset. DenseNet121 achieved the most balanced performance (Accuracy = 96.2%, Recall = 95.5%, F1-score = 96.2%), prioritizing the reduction of false positives and negatives. The model was integrated into a mobile application developed under the Extreme Programming methodology, employing a client-server architecture with RESTful APIs. The NLP component, based on the DeepSeek-R1 model through the OpenRouter platform, generated contextually relevant clinically responses. Validation with healthcare professionals showed that 91.7% would recommend it as a support tool for tuberculosis diagnosis, due to its diagnostic reliability, usability, and clinical recommendations. These findings confirm the clinical viability and practical utility of the application, positioning it as a valuable resource in healthcare contexts with limited specialist availability