Abstract

Introduction: One of the problems affecting health in Chile refers to cerebral pathologies, the performance of tests and the long wait to obtain the results (delays in diagnosis and treatment). Currently, the exams are sent abroad to be processed and the waiting time plays against the patient. Given this reality, our paper proposes a deep learning model for brain image prediction that allows obtaining an early, but not definitive, diagnosis in order to decrease the process time and, if necessary, prioritize patients whose life would be potentially at risk. Methods: The development used an iterative DAR approach and the images were collected from Kaggle. In addition, the dataset is resized to normalize the size and we generated new images using "data augmentation". The images were processed on convolutional networks, investigating different configurations of the network, its optimizer and the activation function, until we arrived at a model that we consider reasonable. Results: With the definitive model, the results exceed 80% hits in the predictions and we found that separating pathologies (hemorrhages and tumors) were fundamental for this result. Conclusions: We have achieved an early diagnostic tool, but the research must be continued because of the increased accuracy. A next step is to expand the dataset with images from other sources and separate the model to analyze pathologies independently. We encourage further research as this type of support can help save lives.