In vivo dynamic monitoring of circulating melanoma cells and the inhibitory effect of PD-L1 inhibitor based on PAFC equipped with a deep learning framework

Abstract

Melanoma is a highly metastatic and lethal skin tumor originating from melanocyte malignancy. Circulating tumor cells (CTCs) are key endogenous biomarkers in melanoma metastasis. Melanin and blood vessels exhibit substantial disparities in their absorbance profiles at select wavelengths, a characteristic that can be adeptly harnessed to differentiate the photoacoustic signals they generate. Photoacoustic flow cytometry (PAFC), which harnesses this principle, enables the monitoring of CTC flowing in vivo. However, this technique is constrained by the inefficiency and high false positive rates associated with traditional algorithms, including the Pauta criterion. In this study, a PAFC system is developed to identify dynamic features of flowing CTCs and the inhibitory effects of PD-L1 inhibitors, using a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory (LSTM) network. The 1D-CNN achieves a balance between classification accuracy and speed. Meanwhile, the LSTM exhibits superior specificity but limited sensitivity. By combining the advantages of the two networks, the inhibitory effect of PD-L1 inhibitors that reduce the CTCs in the blood and block metastasis to other organs of melanoma mouse models are studied noninvasively in vivo and validated in vitro. The PAFC equipped with the deep learning framework provides a more timely and efficient assessment of PD-L1 inhibitors compared to conventional pathological methods, significantly enhancing the melanoma diagnosis and treatment monitoring. This technology demonstrates potential as a significant tool for the non-invasive, dynamic evaluation of melanoma progression and response to immunotherapy.