Explainable Artificial Intelligence in Malignant Lymphoma Classification: Optimized DenseNet121 Deep Learning Approach With Particle Swarm Optimization and Genetic Algorithm

• Published in: IEEE access Vol. 13; pp. 98639 - 98655
• Main Authors: ELwahsh, Haitham, Bakhiet, Ali, Alirr, Omar Ibrahim, Khalifa, Tarek, Alsabaan, Maazen, Ibrahem, Mohamed I., El-Shafeiy, Engy
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Artificial intelligence; Artificial intelligent; Artificial neural networks; Cancer; Classification; Computational modeling; Deep learning; Explainable AI; Explainable artificial intelligence; Feature extraction; Freezing; genetic algorithm; Genetic algorithms; Lymphoma; Machine learning; malignant lymphoma; Medical diagnostic imaging; Medical imaging; Optimization; optimized DenseNet121; Particle swarm optimization; Swarm intelligence
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3575364

One of the forms of cancerous tumors that can be fatal is malignant lymphoma. Histopathological examination of lymphoma tissue images is a diagnostic technique for detecting malignant lymphomas. Differentiating lymphoma subtypes manually is challenging due to their similar morphological features. This paper introduces a cutting-edge methodology for classifying malignant lymphoma using Hematoxylin and Eosin (H&E)-stained biopsy images. It integrates feature reduction using Particle Swarm Optimization (PSO) and model optimization through a Genetic Algorithm (GA) in conjunction with DenseNet121 architecture. The DenseNet121 model, which employs four distinct freezing strategies, extracts features from the images, generating an initial 4096 features per image. These features are subsequently reduced by more than 54.4% using PSO, thereby enhancing the computational efficiency without sacrificing classification accuracy. The trained GA Deep Neural Network (DNN) model achieved a significant score of 96.77% testing accuracy with high precision, recall and F1 scores as well. The results after comparison with other existing state-of-the-art models showed the usefulness of the proposed approach in terms of performance as well as the consumption of resources for medical imaging classification. It is recommended that the combined application of PSO for feature reduction and GA for model optimization can be successfully used for improving accuracy rate of such algorithms while reducing computation time. This technique is expected to be further developed and applied in different tasks of computer aided diagnosis of malignant lymphoma in particular the tasks with requirements for explainable artificial intelligence.