Optimized deep learning ensemble using Fast Osprey algorithm for accurate lymphoblastic leukemia detection

Introduction Acute Lymphoblastic Leukemia (ALL) is a hematological malignancy, which is life-threatening and demands rapid and precise diagnosis to either enhance or worsen the survival chances. Traditional diagnostic methods, especially the manual microscopic examination, are labor-intensive and subject to inter-observer variability. Even though deep learning models have been shown to achieve good performance in automated detection, single-model structures tend to be prone to overfitting and under-generalize to heterogeneous datasets, with low interpretability. Hence, an effective and responsive computer-aided diagnostic (CAD) platform is required to promote the reliability of diagnostics. Methods We introduce a new ensemble-based model that can be trained on a combination of several state-of-the-art convolutional neural networks (CNNs), such as EfficientNetB3, EfficientNetV2B3, and EfficientNetV2B1, and optimized with Fast Osprey Optimization (FOO), a bio-inspired algorithm that dynamically assigns optimal ensemble weights. An extensive dataset was formed through the combination of all publicly available datasets, and thereafter, data augmentation was used to address the issue of class imbalance and to improve the generalization of the model. The FOO algorithm is a model contribution optimization algorithm that is used in the training process to enhance predictive robustness and computational efficiency. Results The proposed FOO-Ensemble model outperformed all baseline architectures. It achieved an accuracy of 97.76%, a precision of 98.13%, a recall of 97.71%, and an F1-score of 97.83%. In addition to improved classification performance, the ensemble approach reduced inference time compared to individual models. Comparative analysis with recent state-of-the-art methods further demonstrated the robustness, scalability, and superior generalization capability of the proposed framework. Conclusion The results demonstrate the usefulness of using deep learning ensembles with bio-inspired optimization in trustworthy ALL detection. A dynamic weighting mechanism improves stability and minimizes the risks of overfitting of standalone models. The higher diagnostic quality and computational capability have high chances of real clinical application. The suggested FOO-Ensemble framework is a scalable and reliable CAD model that will be able to assist hematopathologists in making early and accurate diagnoses of ALL, which will ultimately result in the provision of better patient outcomes.