Deep learning-based cancer cell classification using morphological and topological organelle features

In the eld of cancer research, there has been a growing interest in understanding how variations in cell morphology contribute to cancer cell dynamics and phenotypes. Across dierent breast cancer cell lines, extensive phenotypic and morphologic heterogeneity has been observed primarily through microscopy. Various image analysis techniques have been employed to discern prominent features associated with specific cancer cells. Machine learning and deep learning algorithms have played a pivotal role in facilitating this feature extraction process. However, extracting alternative yet valuable distinctive features is challenging for cell line classification due to physical and technical limitations with adequate data acquisition. To address this challenge, we developed a workflow that utilizes fine-tuned patch-based image preprocessing and self-supervised learning for deep learning-based classification of cancer cells within confocal microscopy images. First, image preprocessing is employed by image partitioning and sparsity filtering of the entire 3D microscopy image into non-sparse 2D image patches. Following the generation of an image patch dataset, a convolutional neural network (CNN) pretrained from self-supervised learning (SSL) tasks is employed for breast cancer cell classification based on features specifically related to organelles. Our proposed methodology outperforms conventional deep learning approaches when tested on a dataset comprising various breast cancer cells. These findings establish a foundation for the robust classification of cancer cells while offering the potential to identify informative organelle characteristics.