Comparative Analysis of ConvNeXt and EfficientNet-B0 for Early Leukemia Detection through Blood Cell Classification with Grad-CAM Interpretability

Leukemia is a hematological malignancy requiring early and accurate diagnosis for optimal patient outcomes, yet conventional microscopic examination remains subjective, time-consuming, and prone to inter-observer variability. This study presents a comprehensive comparative analysis of two state-of-the-art deep learning architectures EfficientNet-B0 and ConvNeXt-Tiny for multi-class blood cell classification aimed at early leukemia detection. Using a balanced dataset of 5,000 microscopic images encompassing five clinically significant classes (basophil, erythroblast, monocyte, myeloblast, and segmented neutrophil), both models were trained and evaluated under identical configurations with extensive data augmentation. Performance assessment encompassed classification metrics, inference speed, and interpretability through Gradient-weighted Class Activation Mapping (Grad-CAM) validated by randomization and occlusion tests. Results demonstrated that both architectures achieved exceptional performance with F1-scores exceeding 98% (EfficientNet-B0: 0.9893, ConvNeXt: 0.9920). ConvNeXt exhibited superior accuracy in distinguishing morphologically similar cells, attributed to its larger receptive fields and advanced architectural design, while EfficientNet-B0 demonstrated dramatic computational advantages with 134 FPS throughput and a compact model size of 18.3 MB six times smaller than ConvNeXt. Grad-CAM visualizations confirmed that both models focus on clinically relevant features including nuclear morphology and cytoplasmic characteristics, validated by low correlation with randomized models (average correlation