Efficient Segmentation and Classification of Leukemia Cells via Optimized Deep Graph Attention Network

Abstract

Leukemia is a cancer affecting blood-forming tissues, leading to an abnormal increase in white blood cells (WBCs), making early and accurate diagnosis essential for improving treatment outcomes and patient prognosis. Traditional diagnostic methods such as microscopic examination of blood smears are time-consuming, subjective, and reliant on expert interpretation. Similarly, conventional machine learning techniques struggle with misclassification, low detection rates, and high computational costs. To address these challenges, this study proposes a novel and efficient deep learning-based classification framework for leukemia detection. The approach begins with the collection of input images from publicly available datasets, which are pre-processed using the Pre-Gaussian Discrete Wavelet Transformer (PGDWT) to enhance image quality. Segmentation of the cancerous regions is performed using the Anatomy-Aware Hover Transformer (AAHT), followed by feature extraction. Classification is then executed using a deep Similarity-Navigated Graph Diffusion Kernel Attention Network (SNGDKANet), specifically designed for precise categorization of various leukemia types. To further improve classification accuracy, the Osprey Optimization Algorithm (OOA) is employed for hyperparameter tuning. The model’s performance is rigorously evaluated based on several metrics, including accuracy (99.5%), sensitivity (99.8%), recall (99.4%), specificity (99.7%), F1-score (99.6%), and precision (99.7%), with a reduced error rate of 3% and an execution time of just 7 seconds. This performance is benchmarked using three well-known leukemia datasets: SN-AM, ALL-IDB, and C-NMC 2019. The results affirm that the proposed framework significantly outperforms existing methods, offering a highly reliable, fast, and accurate tool for leukemia diagnosis in clinical and research settings.