Machine learning-supported identification of necrosis in colorectal and pancreatic cancer spheroids.

Abstract

The transition to three-dimensional cell models is a reality in cancer research. While these models are more physiologically relevant, they present challenges that have not been previously considered. Certain aspects of 3D cultures, such as cell death, need to be modelled as they have a profound impact on spheroid survival and therefore on cancer treatment testing. To identify necrosis in HT-29 and Hs766-T cancer spheroids using an accurate, cost-effective, and fast machine learning-based system that relies only on brightfield (BF) images. In this study, three staining procedures with dyes were performed in colorectal and pancreatic cancer spheroids to highlight necrosis. The microscopic images of these stains serve as a ground truth for different machine learning models, which can subsequently learn to detect necrosis from the brightfield image, i.e., without the need for staining. The results show a strong correlation between dark spots in the spheroid and their staining as necrotic, making regression models an appropriate solution. In general, the supervised models performed better than the unsupervised ones. The models perform worse in all cases for pancreatic Hs-766T spheroids, where necrosis is manifested as small, scattered spots, as opposed to colorectal HT-29 spheroids, where necrosis is a larger cluster in the center. The U-Net regression models demonstrated particularly noteworthy performance, at times surpassing 90% Normalized Coefficient of Correlation. The study findings demonstrate the potential for modeling necrosis in spheroids to a certain extent using 2D microscopy images and machine learning identification. The accuracy of the models largely depends on the manifestation of the necrosis, that is, the specific cancer cells that make up the spheroids.