Detection of retinitis pigmentosa stages with GAN and transfer learning in maps of MfERG P1 wave amplitudes

Retinitis pigmentosa (RP) is an inherited disease caused by the dysfunction of photoreceptor cells in the retina. Effective management and therapy are crucial to delaying or preventing vision loss and blindness, making RP a leading cause of inherited retinal disorders globally. This study is the first to utilize maps of P1 wave amplitudes from multifocal electroretinography for the automatic identification and staging of RP. We enhanced raw data with image preprocessing, traditional image augmentation, and generative adversarial networks (GAN) to create comprehensive datasets, followed by classification using five transfer learning models. Initial results indicated overfitting, mitigated through advanced preprocessing and augmentation. Performance metrics showed that the Deep Convolutional GAN model effectively eliminated overfitting. ResNet50 achieved the highest performance in four-class classification with 95.2% recall, 95.2% precision, 94.9% accuracy, 98.2% specificity and 95.2% F1 score. These results suggest that maps of P1 wave amplitudes are valuable indicators for differentiating RP stages, highlighting their potential for improving RP diagnosis and management.