VNLU-Net: Visual Network with Lightweight Union-net for Acute Myeloid Leukemia Detection on Heterogeneous Dataset

Abstract

Recent advancements in Artificial Intelligence (AI) and Deep Learning (DL) have shown promising results in Acute Myeloid Leukemia (AML) detection. However, challenges remain due to limited, annotated datasets and the need for precise architectures. This paper proposes VNLU-Net, a novel DL framework, by integrating the frozen VGG16 with a lightweight Union-net module. The Union-net module substitutes the last three convolutional layers along with fully connected layers of VGG16. In the framework, the frozen layers of VGG16 provide robust feature extraction, leveraging the pretrained weights. Subsequently, the Union-net refines these features with minimal parameters, enhancing model robustness and generalization. The proposed method achieves better performance, with 99.37% accuracy on the BBCI_AML_2024 dataset and 99.71% on a heterogeneous dataset. Additionally, the qualitative analysis using Gradient-weighted Class Activation Mapping (Grad-CAM) establishes the efficacy of the model. Moreover, the comparative analysis signifies its superiority over the standard existing approaches in terms of accuracy, precision, recall, F1-score, and specificity.