Comparativa de redes de aprendizaje profundo empleando transfer learning para detección de cáncer de mama en el corpus CBIS-DDSM

Abstract

The objective of this study is to compare various deep learning network architectures using transfer learning for breast cancer detection in images from the CBIS-DDSM corpus. Models such as VGG16, ResNet50, and EfficientNet were evaluated, trained, and fine-tuned with preprocessed mammographic images. The results show that the ResNet152 model achieves the best performance in terms of accuracy and sensitivity, standing out as a promising tool for supporting the early diagnosis of breast cancer. Additionally, the study includes the application of image processing techniques such as contrast enhancement using CLAHE and morphological operations. The results indicate that processed mammograms show superior performance. The ResNet50 network enhanced with CLAHE achieved an accuracy of 68.2%, while with morphological operations it achieved 67.7%. However, the VGG19 network with normal mammograms showed the best performance, with an accuracy of 69.1%.