EAL-Net: Evolutionary Algorithm-optimized Novel Lightweight Deep Learning Architecture with Multi-attention Modules and Orthogonal Softmax for Blood Cell Diagnosis

Traditional diagnostic methods for Blood Cells (BCs) and Cervical Cells tend to erroneous analysis and inaccurate diagnoses due to the complexity involved in overlapping regions of cells. Deep Learning (DL) techniques through Convolutional Neural Networks (CNNs) have gained prominence in addressing these issues but encounter challenges as inefficient feature attention, and high computational processing power. To tackle these inefficiencies, this paper presents a novel computationally efficient lightweight architecture (EAL-Net) for BCs diagnosis, which is based on lightweight multi-attention modules and Orthogonal SoftMax Layer (OSL). OSL functionality precisely mitigates parameter co-adaptation to ensure orthogonality among weight vectors, thus refining feature learning processes. The developed EAL-Net architecture is equipped with novel Lightweight Attention Mechanism (LWAM) to selectively learn discriminative features and enabling more precise attention on diagnostic relevant regions. LWAM integrates Spatial-Attention Module (SAM) to focus on key regions, Spatial Self-Attention Module (SSAM) to capture long-range dependencies, and Category-Attention Module (CAM) to emphasize class-specific features. Collectively, this approach amplifies feature attention and strengthening spatial and channel-wise dependencies, significantly improving the performance of neural network architecture. Moreover, Genetic Algorithms (GA) efficiently search the optimal set of hyperparameters iteratively. GA is chosen for its efficacy in handling complex search spaces and non-linear optimization problems. Additionally, Grad-CAM and SHAP provide model insights by highlighting key features influencing predictions. EAL-Net outperforms pre-trained CNNs, transformer-based architectures, and existing methods across accuracy, efficiency, and memory use. With only 0.73 million parameters and 8.2 MB size, it achieves 96.11% accuracy on 8-class and 93.68% on 19-class BCs datasets. To validate the robustness of the proposed architecture across diverse medical imaging modalities, we evaluated EAL-Net on Cervical Cells (SIPaKMeD & Mendeley_LBC), histopathological and chest X-ray datasets.