Investigação de técnicas de normalização e segmentação com características interpretáveis para imagens histológicas H&E

Abstract

Oral epithelial dysplasia is a precancerous condition that can be identified through the analysis of histological images, typically performed by specialists. However, the high workload and variability in clinical experience make this process susceptible to subjectivity. As a form of support, computational tools have been used as a supplementary aid in decision-making. This work presents an investigation into different Hematoxylin and Eosin (H&E) stain normalization methods in histological images and the development of a nuclei segmenter based on a U-Net architecture with feature interpretability. Dictionary learning-based stain normalization methods were tested and evaluated using four different metrics. These normalized datasets were assessed using a modified U-Net segmentation architecture. Interpretability of the descriptors in the internal layers of the model was explored to evaluate the influence of normalization and the enhancement of segmentation performance through the application of the ProtoSeg technique. The results showed that the normalization method based on Euclidean Distance yielded average values of 0.897, 44.401, 0.898, and 0.973 for the FSIM, PSNR, QSSIM, and SSIM metrics, respectively. Furthermore, the findings indicated that certain normalization methods improved the segmentation stage, reaching Dice coefficient values around 0.7683. However, when the ProtoSeg strategy was applied alongside normalization, these results were not superior. This study demonstrated that H&E stain normalization had a positive impact on segmentation performance only when interpretability was not applied during the segmentation process.