Sample Viability Assessment from H &E Whole Slide Images for Lung Cancer Molecular Testing
Autores: Ruth Román Cristina Aguado Josep Castellvi Sonia Rodríguez Erika Aldeguer María Teresa Torrijos
Fecha: 14.06.2023
Abstract
Molecular testing has become an essential tool in precision oncology as targeted therapies have shown to increase the survival rate of patients. However, for molecular test results to be reliable, some requirements must be met, such as the presence of a minimum percentage of tumor cells in a minimum area of the sample. Currently, this analysis is performed by viewing the histopathological slides under the microscope and manually quantifying and highlighting the areas with the highest tumor cellularity. This results in low reproducibility and high subjectivity. To address these problems, we propose a deep learning framework to assist pathologists in the evaluation of the viability of a sample for molecular testing. The developed approach highlights viable sample regions, as well as areas that require further processing. To this aim, we implement a 3-step methodology to analyze Whole Slide Images (WSI): 1) stain normalization of WSI tiles, 2) classification of tiles by a cascade approach, 3) heatmap generation to determine the area of the WSI to perform molecular testing. Moreover, we use three lung cancer subtypes and compare the performance when the models are trained separately for each type or jointly. We achieve a F1-score of 0.63 at tile-level, while at the WSI-level the F1-scores were 0.71 and 0.96.
BIB_text
title = {Sample Viability Assessment from H &E Whole Slide Images for Lung Cancer Molecular Testing},
pages = {365-374},
keywds = {
Biological organs; Classification (of information); Deep learning; Diseases
}
abstract = {
Molecular testing has become an essential tool in precision oncology as targeted therapies have shown to increase the survival rate of patients. However, for molecular test results to be reliable, some requirements must be met, such as the presence of a minimum percentage of tumor cells in a minimum area of the sample. Currently, this analysis is performed by viewing the histopathological slides under the microscope and manually quantifying and highlighting the areas with the highest tumor cellularity. This results in low reproducibility and high subjectivity. To address these problems, we propose a deep learning framework to assist pathologists in the evaluation of the viability of a sample for molecular testing. The developed approach highlights viable sample regions, as well as areas that require further processing. To this aim, we implement a 3-step methodology to analyze Whole Slide Images (WSI): 1) stain normalization of WSI tiles, 2) classification of tiles by a cascade approach, 3) heatmap generation to determine the area of the WSI to perform molecular testing. Moreover, we use three lung cancer subtypes and compare the performance when the models are trained separately for each type or jointly. We achieve a F1-score of 0.63 at tile-level, while at the WSI-level the F1-scores were 0.71 and 0.96.
}
isbn = {978-981993310-5},
date = {2023-06-14},
}
