Deep Transfer Learning and Ensemble Learning Coupled with Non-Linear Optical Microscopy for the Classification of Senescent Cells

Recent oncology research highlights that senescence, once deemed beneficial in cancer treatments, can contribute to cancer relapse. Detecting therapy-induced senescent cells is challenging due to their complexity and lack of specific markers. Nonlinear optical (NLO) microscopy provides a fast, non-invasive, label-free detection solution. To distinguish between senescent and proliferating cells, here we present the development of a deep learning architecture based on multimodal NLO microscopy images coming from Stimulated Raman Scattering, Two Photon Excited Fluorescence and Optical Transmission. Despite limited labeled data, Transfer Learning, Data Augmentation, and Ensemble Learning techniques allowed us to achieve an accuracy over 90%. Ultimately, the predictions of the neural network are evaluated using the Grad-CAM visualization approach, which allows highlighting the most important features in the input images responsible for the labels assigned by the network. This work reveals the effectiveness of deep learning in senescence classification, potentially advancing treatment strategies.