An innovative approach to parasite classification in biomedical imaging using neural networks

Abstract

Background and objective: In light of the increasing demand for precise and effective parasite detection in biomedical imaging, current models frequently depend on widely used convolutional neural networks (CNNs), with minimal innovation in feature selection and engineering for artificial intelligence. This study addresses these gaps by introducing a novel CNN, termed BLGSNet (Batch Normalization, Layer Normalization, GELU - Gaussian Error Linear Unit - and Swish functions-based network), specifically designed for parasite detection. BLGSNet increases feature engineering deploying a unique intersection-based feature selection method integrated into the deep learning model to improve classification accuracy. Materials and methods: Using a publicly accessible image dataset with eight classes, including two blood cell types and six parasite types, we developed the BLGSNet architecture, which incorporates transformer-inspired design, convolution-based residual blocks, batch normalization, layer normalization, GELU, and Swish activation functions. The accompanying deep feature engineering model employs transfer learning and includes four key phases: (i) feature extraction using BLGSNet, (ii) feature selection using intersection methodology with neighborhood component analysis, chi-square, minimum redundancy maximum relevance, and ReliefF feature selectors, (iii) classification with k-nearest neighbors (kNN) and support vector machine (SVM), and (iv) automated selection of the optimal output. Results: The proposed BLGSNet was trained on the designated dataset, achieving a test accuracy of 99.25%, while the deep feature engineering model attained an improved test accuracy of 99.59%, demonstrating the model's robustness. Conclusion: The high classification accuracy achieved by BLGSNet and the feature engineering model highlights their potential in addressing the complex task of parasite detection.