Deep learning-assisted classification of phototherapy-induced cellular responses mediated by gold shell-isolated nanoparticles.

Photodynamic therapy (PDT) and photothermal therapy (PTT) can induce cellular responses that are reflected not only in viability assays but also in subtle morphological and textural changes in post-treatment cell images. Here, we investigate whether label-free optical microscopy images can be used to classify treatment-associated phenotypes in melanoma cells exposed to photoactive nanostructures. Primary A375 and metastatic SH-4 melanoma cells were treated with Methylene Violet 3RAX (MV), gold shell-isolated nanoparticles (AuSHINs), or the combined AuSHINs@MV@ system under irradiated and non-irradiated conditions. Deep convolutional neural networks were applied to classify images according to treatment and concentration, while MTT assays performed under the same experimental conditions were used as an independent biological reference for cell viability. The MTT results showed strong phototoxicity for MV and AuSHINs@MV@ under irradiation, whereas AuSHINs alone produced limited effects in melanoma cells. Image-based classification distinguished treatment- and concentration-associated patterns with an average accuracy of approximately 93%, with higher performance generally observed for irradiated groups. Gradient-weighted Class Activation Mapping (Grad-CAM) analyses further indicated that the models relied on textural and spatial features not readily evident by visual inspection. These findings support the use of machine learning-assisted optical microscopy as a complementary, low-cost screening strategy to identify image phenotypes associated with phototherapy-induced cellular responses and to guide the selection of experimental conditions for further biological validation.