Computational method of spectral matching for color normalization of H&E histological images by autoencoders

• Grant number: 23/15834-6
• Support Opportunities: Scholarships in Brazil - Master
• Start date: January 01, 2024
• End date: December 31, 2025
• Field of knowledge: Engineering - Biomedical Engineering - Medical Engineering
• Principal Investigator: Thaína Aparecida Azevedo Tosta
• Grantee: Hanna Beatriz Couto Monteiro Fernandes de Castro
• 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

Diagnosis of different types of cancer can be confirmed by analysis of images obtained by tissue samples digitization stained by Hematoxylin-Eosin (H&E). These histological images can be commonly used by scientific research for the proposal of computer systems that aid pathologists. However, the performance of these systems can suffer from the influence of color variations of H&E histological images. To deal with these variations, a normalization technique can be used to color-adjust these images. Recent works have presented new normalization methods but with limitations, such as the absence of biological concepts and spatial information in their methodological definitions. Therefore, this project proposes a new approach for H&E histological images normalization by spectral matching and autoencoders techniques. These concepts allow stain colors to be estimated by dictionary learning techniques based on biological concepts and use structural information of the images for this purpose. This proposal also integrates the investigation of computational techniques for initializing these estimates and the unsupervised definition of the parameter for the representation of stain sparsity by bioinspired algorithms. For performance evaluation, it is expected that public histological images are used, with variations of contrast, color, and magnification. To do so, visual evaluations will be performed as well as quantitative analyses of normalized image quality and of its use in computer-aided diagnosis systems.