Ocotillo optimization-driven deep learning for bone marrow cytology classification

Abstract

Manual diagnosis of hematological cancers like leukemia through bone marrow smear analysis is labor-intensive, prone to errors, and highly dependent on expert knowledge. To overcome these limitations, this study introduces a comprehensive deep learning framework enhanced with the innovative bio-inspired Ocotillo Optimization Algorithm (OcOA), designed to improve the accuracy and efficiency of bone marrow cell classification. The contributions include developing a baseline Convolutional Neural Network (CNN) that achieves an initial accuracy of 86.29%, surpassing existing state-of-the-art deep learning models. Further integrate the binary variant of OcOA (bOcOA) for effective feature selection, which reduces the average classification error to 0.4237 and increases CNN accuracy to 93.48%. Additionally, utilize the continuous version of OcOA for hyperparameter optimization, further enhancing CNN performance to a maximum accuracy of 98.24%. Crucially, this optimization also results in a substantial clinical performance gain, with sensitivity increasing from 86.02% to 98.34% (+12.32%), specificity rising from 86.53% to 98.14% (+11.61%), and the false negative rate being significantly reduced, thereby enhancing diagnostic reliability in critical scenarios. These findings highlight the potential of metaheuristic optimization techniques to improve the effectiveness of deep learning models in clinical diagnostics quantifiably. The proposed approach demonstrates measurable gains in automated cytology technology, offering a scalable, interpretable, and accurate solution for hematological screening applications.