The increasing prevalence of white blood cell (WBC) diseases, such as leukemia and lymphoma, presents a significant challenge in the healthcare field. Effective treatment and improved patient outcomes are crucial for early and accurate detection. This study investigated the potential of an advanced automated detection system based on deep learning algorithms to address the critical issue. It evaluated the use of EfficientNetB0, ResNet50, and convolutional neural network - multilayer perceptron (CNN-MLP) models for the detection of WBC diseases. The experiments were carried out using a dataset split into 60% training, 20% validation and 20% testing. EfficientNetB0’s high accuracy in classifying WBCs, along with its excellent precision and recall, is crucial for minimizing false positives and negatives in diagnoses. Furthermore, the study examines the impact of data preparation approaches such as resizing, augmentation, and normalization on model performance and generalizability. The results showed that all models achieved over 90% accuracy, with EfficientNetB0 attaining 99.55% using data augmentation and ResNet50 achieving the highest accuracy of 99.70% under Random Oversampling. In contrast, CNN-MLP consistently underperformed relative to the other models. Oversampling techniques were employed to ensure that the dataset had a complete representation of all types of white blood cells. The findings revealed that deep learning can significantly improve the detection of WBC diseases, leading to more efficient and reliable diagnostic systems in healthcare. This research highlights the pivotal function of artificial intelligence in enhancing diagnostic precision and facilitating optimal clinical decision-making. EfficientNetB0 stood out with its high precision, recall, and F1-score, effectively reducing diagnostic errors and proving more suitable for clinical applications. The study demonstrates that the integration of EfficientNetB0 with data preparation techniques significantly enhances model generalizability and diagnostic reliability.
Indra et al. (Wed,) studied this question.