Class Weighting in Imbalanced Data for Leukocyte Classification using Fine-Tuning Inception-v3

This study investigates the classification of leukocyte images in an imbalanced dataset using deep learning techniques. The dataset consists of 14,514 images, categorized into five leukocyte types: basophils (301), neutrophils (8,891), lymphocytes (3,461), monocytes (795), and eosinophils (1,066). To address class imbalance, we applied class weighting alongside transfer learning and fine-tuning using the Inception-v3 architecture. The dataset was split into 80% for training and 20% for testing, and 5-fold cross-validation was conducted to evaluate model robustness. Hyperparameters were set with a learning rate of 0.0001, batch size of 32, and 30 training epochs, optimized using the Adam optimizer. Fine-tuning was performed on selected layers, specifically the Inception modules Mixed_6d and Mixed_7a, which showed significant performance improvements. The results demonstrated an accuracy of 97.26%, precision of 97.34%, recall of 97.26%, and an F1-score of 97.28%. This indicates that combining class weighting with fine-tuning on deeper layers leads to enhanced model performance, particularly for minority classes in imbalanced datasets. Our findings suggest that this approach can be effectively applied to improve the classification of medical images in tasks involving class imbalance.