Comparative evaluation of CNN and SOM in the detection of tuberculosis in chest radiographs using a mobile platform

Resumen

This research presents the development of an automated diagnostic system for the early detection of tuberculosis (TB) in chest radiographs by comparing convolutional neural networks (CNN) and self-organizing maps (SOM), in order to evaluate which one offered better results in medical classification tasks. For the CNN model, pre-trained MobileNetV2 was used as feature extractor being fine-tuned with augmented and normalized images, achieving 98 % accuracy in binary classification (normal or tuberculosis). On the other hand, the SOM model was trained with the vectors generated by the CNN and allowed visualizing the distribution of the data in the feature space, achieving an accuracy of 95 %. Therefore, after comparing the performance of both models, it was chosen to implement the CNN as the core of the system, due to its high accuracy and generalization capability. The final model was integrated into a mobile application that connects to cloud services using Hugging Face for inference and Firestore for results storage. This solution was especially designed for resource-limited contexts, allowing health professionals in rural or hard-to-reach areas to have a support tool for the preliminary diagnosis of TB from a smartphone.