Multiclass classification of ocular toxoplasmosis through transfer learning with convolutional neural networks

Ocular toxoplasmosis (OT) is an infection caused by the parasite Toxoplasma gondii, which can lead to serious ocular damage and vision loss. Accurate diagnosis of OT is essential for effective treatment; however, it remains challenging due to its wide range of clinical manifestations. We propose a deep learning (DL) and transfer learning-based approach to improve OT diagnosis using fundus images. Five pre-trained models were evaluated: DenseNet201, ResNet18, VGG16, MobileNetV2 and InceptionV3. A total of 412 fundus images were used, labeled as healthy or presenting active or inactive OT lesions. Data augmentation, Optuna-based hyperparameter optimisation, and cross-validation were applied. DenseNet201 achieved the best overall performance, with a recall of 95.09% and a specificity of 97.39%. The remaining models also achieved strong results in terms of specificity. Overall, transfer learning in convolutional neural networks coupled with hyperparameter optimisation proved to be highly effective for OT image classification.