Comparative analysis of stained normalization in H&E histopathological images of breast cancer for nuclei segmentation improvement

• Main Authors: Reyes, Hector Eduardo Zepeda, Barreto, Hayde Peregrina, Ramos, Jose Alfonso Cruz, Del Carmen Lopez Armas, Gabriela
• Format: Conference Proceeding
• Language: English
• Subjects: Breast cancer; Color Normalization; Histopathology; Image color analysis; Image segmentation; Neural Networks; Nuclear Segmentation; Task analysis; Visual Breast Cancer Detection; Visual databases; Visual Methods in Histopathology Diagnosis; Visualization
• ISSN: 2573-0770
• DOI: 10.1109/ROPEC58757.2023.10409338

According to the World Health Organization, breast cancer is defined as the abnormal and disorganized growth of cells in breast tissue. It is currently one of the biggest challenges facing health systems worldwide. Making a precise diagnosis is essential to offer the most appropriate treatment adapted to each patient's condition. Computer vision-based tools could help the specialist in the diagnosis task. One major challenge in digital histopathology is the variation from the staining in the tissue. This work assesses the performance of color normalization under different approaches, measuring its impact on the structures of interest related to the graduation of tissue samples. It also analyzed how color normalization could improve relevant structure segmentation, particularly nuclei, an essential element in the Nottingham Graduation System followed by pathologists. Four methods of color normalization were implemented, and their respective structural and color-associated metrics (SSIM, PSNR, and colorfulness) were calculated. With the Reinhard method, the benign group got the highest SSIM 0.9744 ± 0.1417. For the normal group, the highest value in SSIM was obtained from the Stain-Net method 0.9403 ± 0.0109. Finally, the highest SSIM value we obtained for the invasive group was 0.9388 ± 0.0090 from the Stain-Net method. After the color normalization methodology, we got a segmentation using the StarDist tool. We can generally observe better segmentation using the color normalization methods proposed by Reinhard and Stain-Net.