Enhancing Leukemia Diagnosis through Machine Learning: A Customized ResNet50 Approach

Leukemia is a type of blood and bone marrow cancer that continues to pose a major challenge in the field of healthcare. This is predominantly caused by the wide range of leukemia subtypes and the complexity of the classification process. Therefore, the current research aims to propose an approach to classify leukemia – more specifically, acute lymphoblastic leukemia – using a comprehensive methodology that integrates various machine learning techniques. A thorough data preprocessing is conducted and utilized Otsu's Thresholding for image segmentation and ResNet50 convolutional neural network for feature extraction. By applying the Lazy Predict library, the proposed work is further streamlined, allowing to finalize the offline model selection process. Finally, a custom ResNet50 CNN model is designed with extra layers in the feature extraction and classification resulting better performance compared to other state of the methods with an accuracy of 0.87 and a balanced accuracy of 0.83. Besides, the proposed model has faster inference times, making its application feasible for real-time leukemia diagnosis.