Organelle-based classification of metastatic breast cancer subtypes using self-supervised learning

During metastatic progression, breast cancer cells exhibit behaviors such as organ-specific tropism and varied responses to metastatic inhibitors. However, the similarity in cellular morphology between metastatic subtypes limits the effectiveness of conventional microscopy for their identification. While fluorescence microscopy combined with deep learning enables automated identification of nonvisible features, such models require large annotated datasets. Therefore, limited annotations and inconsistent data handling commonly hinders network generalizability and task performance. To address this, we utilize self-supervised learning (SSL) with sparsity-filtered patch-based preprocessing for user-agnostic classification of metastatic tropism and drug-treated triple-negative MDA-MB-231 variants. A threshold-based pipeline first extracts non-sparse 2D patches from 3D multi-channel confocal images for mitochondria and early endosomes. SSL is then applied using two pretext tasks: a jigsaw task that learns spatial structure by predicting permutations of shuffled sub-patches, and contrastive learning task that learns augmentation invariance by distinguishing similarity between augmented views of the same patch and dissimilarity between unrelated patches. The learned representation is then fine-tuned for two downstream classification tasks: metastatic tropism subtype identification, including parental, lung-tropic, and bone-tropic cells; and characterization of cellular response to varying concentrations of the metastatic inhibitor, MGH-CP1. From these results, we demonstrate that SSL-pretrained models improve classification of metastatic subtypes based on mitochondrial and endosomal distribution, outperforming gold-standard pretraining approaches. This workflow provides a scalable approach for microscopic cancer image analysis and may support future efforts in automated cancer diagnostics. Furthermore, the SSL framework enables generalization to specialized microscopic analysis tasks without requiring labeled datasets, thus making it suitable for broader applications.