A Hybrid Approach for Classification of Lyme Disease using Deep Convolution Neural Networks and Bandelet Transform

The bacterium Borrelia burgdorferi is the source of the complicated tick-borne sickness known as Lyme disease. It is essential to accurately categorize Lyme disease into its many stages and symptoms to assist with treatment choices and enhance patient outcomes. In the proposed work, a unique method for categorizing Lyme illness into 15 different classifications by combining the features of Bandelet transform with Convolutional Neural Networks (CNN) which are employed to learn and categorize the discriminative features that were extracted from medical images. Therefore, multiresolution features are extracted using Bandelet Transform. Statistical Texture features such as Mean, Variance, Standard Deviation, Kurtosis and Skewness are calculated from the coefficients of Bandelet Transform which are concatenated with the features extracted from the different pre-trained networks. For the balanced dataset, using EfficientNet-B0 combined with Bandelet transform achieved an IBA of 0.974 and test accuracy of 0.9868. For the unbalanced dataset, using EfficientNetV2B1 with Bandelet transform resulted in an IBA of 0.932 and test accuracy of 0.9812. When classifying the images into 15 different classes, EfficientNetB7 with Bandelet transform achieved an IBA of 0.6424 and test accuracy of 0.7925. The results demonstrate that concatenating pretrained network features with Bandelet transform features improves classification performance for binary classification tasks (balanced and unbalanced datasets), though performance is more limited when extending to 15-class classification.