Deep features obtained through LIME and Grad-CAM: an analysis exploring H&E normalized images

• Grant number: 25/05775-8
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: June 01, 2025
• End date: December 31, 2025
• Field of knowledge: Engineering - Biomedical Engineering
• Principal Investigator: Thaína Aparecida Azevedo Tosta
• Grantee: Betânia Caroline Silva de Albuquerque
• Host Institution: Instituto de Ciência e Tecnologia (ICT). Universidade Federal de São Paulo (UNIFESP). Campus São José dos Campos. São José dos Campos , SP, Brazil
• Associated research grant: 22/03020-1 - Normalization of H&E stain by autoencoders with analysis of ensemble learning for histological images, AP.PNGP.PI

Abstract

Despite advancements in the field of image analysis, especially through convolutional neural networks (CNN), it is still possible to investigate the discriminative power of hybrid CNN models by considering different forms of image representations, such as those developed to produce visualizations of the neural activation patterns in a CNN. For example, it is possible to use the values from a CNN layer to identify the regions of an image that contribute to a classification, through Grad-CAM and LIME strategies. This allows the evaluation of the discriminative power of H&E histological images after applying different color normalization techniques. This type of investigation has not been fully explored in the literature, especially in defining a model that combines LIME and Grad-CAM image representations. These associations are valuable contributions to the classification and pattern recognition of diseases using H&E images, supporting the development of more complete computational systems. (AU)