Optimized Machine Learning Models for Lymphocytes Image Classification and Leukemia Detection

Leukaemia remains a life-threatening blood cancer that leads to rapid and accurate diagnosis which requires sophisticated methods to improve the outcome of the patient. In this paper, optimized machine learning models for classifying lymphocyte images are explored for leukaemia detection support. A large, diverse dataset of blood smear images with more than 18,000 samples across eight blood cell categories was utilized. The study essentially employs deep learning techniques in the form of ResNet50 with visualization from Grad-CAM to achieve over 99% accuracy when distinguishing leukaemia-related abnormalities. Traditional models, like SVM, Random Forest, and XGBoost are also compared with regards to efficiency and interpretability. ResNet50 demonstrated good performance based on its hierarchical feature extraction and transfer learning capabilities. Interpretable visualization through Grad-CAM was ensured to capture the critical diagnostic features correctly. Experimental results indicate that advanced CNN architectures may improve automated haematology diagnostics when integrated with traditional models. Future avenues could include exploring ensemble methods or other novel architectures such as the Vision Transformer to further perfect it. This work represents a significant step forward in non-invasive, efficient, and accurate leukaemia diagnostics.