Deep Learning-Assisted Microscopy Reveals Progressive Supramolecular Remodeling and Colloidal Reorganization of...

Bovine milk represents a highly complex colloidal system whose physicochemical stability depends on the organization of milk fat globules, casein micelles, membrane-associated phospholipids, and somatic cellular components. Mechanical separation procedures such as centrifugation induce redistribution of dispersed colloidal fractions and structural perturbations within the milk matrix, potentially enabling fraudulent reduction of somatic cell count while preserving bulk compositional parameters. In the present study, we investigated whether advanced deep learning architectures could identify centrifugation-associated structural alterations in bovine milk using microscopy image representations. A total of 16,472 microscopy images obtained from centrifuged and non-centrifuged milk samples were analyzed using Swin Transformer V2 and ConvNeXt-Base architectures. Both models successfully detected centrifugation-associated structural perturbations and substantially outperformed the previously analyzed InceptionC baseline. ConvNeXt-Base achieved 87.30% classification accuracy together with 86.85% balanced accuracy and 86.59% harmonic average of recalls following totalogit aggregation. Importantly, Swin Transformer V2 demonstrated strong monotonic relationships between logit metrics and centrifugation ratio (r = 0.640-0.651, p < 0.01), indicating sensitivity to progressive image-level changes associated with increasing centrifugation ratio. Collectively, the obtained findings demonstrate that microscopy-derived deep learning representations capture structural information associated with centrifugation-induced changes in bovine milk, supporting the applicability of AI-assisted microscopy for detecting processing-related alterations in complex dairy systems.