Deep ensemble learning and transfer learning methods for classification of senescent cells from nonlinear optical microscopy images

The success of chemotherapy and radiotherapy anti-cancer treatments can result in tumor suppression or senescence induction. Senescence was previously considered a favorable therapeutic outcome, until recent advancements in oncology research evidenced senescence as one of the culprits of cancer recurrence. Its detection requires multiple assays, and nonlinear optical (NLO) microscopy provides a solution for fast, non-invasive, and label-free detection of therapy-induced senescent cells. Here, we develop several deep learning architectures to perform binary classification between senescent and proliferating human cancer cells using NLO microscopy images and we compare their performances. As a result of our work, we demonstrate that the most performing approach is the one based on an ensemble classifier, that uses seven different pre-trained classification networks, taken from literature, with the addition of fully connected layers on top of their architectures. This approach achieves a classification accuracy of over 90%, showing the possibility of building an automatic, unbiased senescent cells image classifier starting from multimodal NLO microscopy data. Our results open the way to a deeper investigation of senescence classification via deep learning techniques with a potential application in clinical diagnosis.

Automated summary

The success of chemotherapy and radiotherapy in treating cancer can lead to tumor suppression or senescence induction. However, recent advancements in oncology research have shown that senescence can contribute to cancer recurrence. Nonlinear optical (NLO) microscopy provides a non-invasive and label-free method for detecting therapy-induced senescent cells. In this study, we develop deep learning architectures to classify senescent and proliferating human cancer cells using NLO microscopy images, and compare their performances. The ensemble classifier approach, which uses seven pre-trained classification networks with additional fully connected layers, achieves a classification accuracy of over 90%. This opens the possibility of building an automatic and unbiased senescent cells image classifier based on multimodal NLO microscopy data, with potential applications in clinical diagnosis.