Deep learning-based tool for radiomics analysis of cancer 3D multicellular spheroids

Cancer 3D multicellular spheroids are a fundamental in vitro tool for studying in vivo tumors. Volume is the main feature used for evaluating the drug and treatment effects, but several other features can be estimated even from a simple 2D image. For high-content screening analysis, the bottleneck is the segmentation stage, which is essential for detecting the spheroids in the images and then proceeding to the feature extraction stage for performing radiomic analysis. Thanks to new deep learning models, it is possible to optimize the process for adapting the analysis to big datasets. One of the most promising approaches is the use of convolutional neural networks (CNNs), which have shown remarkable results in various medical imaging applications. By training a CNN on a large dataset of annotated images, it can learn to recognize patterns and features that are relevant for segmenting spheroids in new images. This approach has several advantages, such as manual or semi-automatic segmentation, which are time-consuming and prone to inter-observer variability. Moreover, CNNs can be fine-tuned for specific tasks and can handle different types of data, such as multi-modal or multi-dimensional images. Starting from the first version of Analysis of SPheroids (AnaSP), an open-source software for estimating morphological features of spheroids, we implemented a new module for automatically segmenting brightfield images by exploiting CNNs. In this work, several deep learning segmentation models have been trained and compared using ground truth masks. Then, a module based on an 18-layer deep residual network (ResNet18) was integrated into AnaSP, releasing AnaSP 2.0, a version of the tool optimized for high-content screening analysis.