Deep neural network-based method applied to confocal microscopy images of PDO cheeses

The main goals of this study were to 1) investigate the feasibility of using a confocal scanning laser microscope (CSLM) on Grana Padano (GP) and Parmigiano Reggiano (PR) PDO cheese samples, and 2) explore the potential of applying deep neural network-based method (DNN) to the CLSM images for predicting the PDO type and the age of the cheese, as further tool related to the quality of these labeled cheeses. A total of 18 cheese samples were collected from 6 dairy plants: 3 belonging to GP and 3 to PR PDO chains. For each dairy, 3 ripening times were selected: 12, 20, 36 for GP and 12, 24 and 36 months for PR. A confocal scanning laser microscope (Leica Microsystems CMS GmbH, Germany) operated with Ar/He- Ne laser excited at 488 nm and 559 nm with a 63X oil immersion objective was used for the fat and protein distribution, after staining the samples with Nile Red (fat) and Fast Green FCF (protein). On average, each cheese sample provided 22 images, with a minimum of 15 and a maximum of 45. The round fat droplets (green) within a continuous protein phase (red), were differently distributed between the two PDO cheeses and among ripening times. To explore the potential of applying DNN method to the CLSM images, only green-channel images (N = 130; 69 from GP, 61 from PR) were chosen (Figure 1a). Image data were normalized and augmented before being analyzed in a DNN model with three 2D convolutional layers (relu activation function) followed by MaxPooling and Dropout layers and a final dense layer (sigmoid activation). The learning rate was 5e-4. The DNN model had 85,089 parameters and was run for 200 epochs. In total, 25 images (13 from GP and 12 from PR) were used for validation. The average accuracy was 0.928 in the training set, meaning that the model can learn the underlying patterns and features from the CLSM images to make accurate predictions, whereas the accuracy was 0.844 in the validation set (Figure 1b). This suggests that the model performs good on the data as well. Further steps, such as regularization techniques or increasing the amount of training data, may be necessary to mitigate the overfitting and improve the model's performance on unseen data. The study is part of the Short-Term Scientific Mission included within the WG2 of the CA19145 COST Action: Innovation related to the integration of several NDSS signals for critical issues in food integrity.