Advanced Detection of Acute Lymphoblastic Leukemia Using Integrated Deep Features and Metaheuristic Algorithms

Objectives: Acute lymphoblastic leukemia is one form of blood cancer. This research work suggests the impact of meta-heuristic feature optimization techniques on leukemia diagnosis. Methods: ALL-IDB2 image database is utilized for this purpose which is publicly available. Techniques for pre-processing images include scaling, SMOTE, and augmentation. Segmentation of the nucleus plays a vital role in detecting leukemia. As a novelty, two distinct deep networks, SegNet and ResUNet are deployed for this purpose in place of conventional segmentation techniques. ResUNet performs better than SegNet with a 0.98 accuracy rate. From the segmented images, about 6144 deep features are extracted using the Xception, Inception-v3, and ResNet50 models. The robust features are selected using mRmR and LASSO methods. The features from the LASSO method are taken for further processing based on the ANOVA test results. Utilizing the whale and moth flame optimization techniques, these features are optimized. SVM and Twin-SVM are used as classifiers in the proposed work. The distinctiveness of the work lies in its combination of metaheuristic algorithms and deep feature integration, which improves detection accuracy. Findings: It is found that the optimized features obtained from whale optimization give an accuracy of 0.9828 whereas the obtained accuracy is only about 0.9478 without using the optimization algorithm. Novelty: The integration of metaheuristic optimization and deep features is introduced in the proposed work to enhance leukemia detection accuracy. The suggested work will definitely help clinicians to make more accurate decisions on acute lymphoblastic leukemia diagnosis. Keywords: Inception, Xception, LASSO, Feature concatenation, Whale Optimization, Twin-SVM