COMPARISON OF HYBRID CONVOLUTIONAL NEURAL NETWORKS MODELS FOR DIABETIC FOOT ULCER CLASSIFICATION

In this paper, we present a comparison of four proposed hybrid deep convolutional neural network models for diabetic foot ulcer (DFU) classification to discriminate between abnormal (DFU) and normal (healthy skin) classes. Increasing the depth in single branch deep convolutional neural networks does not always significantly contribute to their overall performance. It may actually lead to a drop in performance due to gradient vanishing issue. Therefore, our proposed models were designed based on the concept of multiple branches network. Traditional convolutional layers and multi-branch parallel convolutional layers were combined to design four deep aggregated models. All four models have six blocks of parallel convolutional layers, but the number of branches of parallel convolutional layers ranges from two to five. Parallel convolutional layers have been employed using different filter sizes on the same input and then concatenated for better feature extraction. To overcome the issues of overfitting and a small amount of training data, we applied several data augmentation techniques. The proposed models were trained with original images first, then with original images plus augmented images, which improved the performance. We empirically prove that a model with four branches outperforms models with two, three, or five branches of parallel convolutions in the task of DFU classification. This model also outperformed the latest DFU classification methods by achieving an F1 score of 95.8% on the DUF dataset.