Structure-based assessment of cancerous mitochondria using deep networks

Mitochondrial functions are essential for cell survival. Pathologic situations, e.g. cancer, can impair mitochondrial function which is frequently reflected by an altered morphology. So far, feature description of mitochondrial structure in cancer remains largely qualitative. In this study, we propose a learning-based approach to quantitatively assess the structure of mitochondria isolated from liver tumor cell lines using convolutional neural network (CNN). Besides achieving a high classification accuracy on isolated mitochondria from healthy tissue and different tumor cell lines which the CNN model was trained on, CNN is also able to classify unseen tumor cell lines, which suggests its superior capability to capture the intrinsic structural transition from healthy to tumor mitochondria.