Análisis comparativo de redes neuronales para la predicción de cáncer de mama en imágenes médicas mediante la evaluación estadística de rendimiento

Abstract

This degree work addresses the comparative evaluation of different convolutional neural network (CNN) architectures for breast cancer prediction in mammographic images. The research focused on identifying the most efficient model by applying metrics such as accuracy, sensitivity and specificity, with the aim of contributing to the development of more accurate and reliable diagnostic tools. To carry out this analysis, data from public databases such as Kaggle and Mini-DDSM were used, which included a representative set of mammograms classified between healthy and cancerous tissue. Pre-trained models such as EfficientNetB0, ResNet50, ResNet152, EfficientNetV2 and ConvNeXt were implemented and optimized using techniques such as hyperparameter fine tuning and class balancing. These adjustments mitigated the challenges associated with data imbalance and variations in image characteristics. The results of the analysis revealed that EfficientNetB0 stood out as the most effective architecture, showing outstanding performance in sensitivity and accuracy, especially when using higher resolution images and dynamic learning rate adjustment strategies. Additionally, the use of preprocessed images and external data increased the generalizability of the models, improving their applicability in clinical scenarios. In conclusion, this study validates the efficacy of convolutional neural networks in breast cancer prediction and highlights the importance of specific optimization strategies to ensure robust performance tailored to the needs of medical practice. The recommendations proposed in this work lay a solid foundation for future projects aimed at integrating these technologies into computer-aided diagnosis systems.