tsta

فارسی

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Robust Binary Differentiation of ALL vs. AML Using Deep Graph Convolutions

Authors :

Mahsan Rahmani¹ Saeed Meshgini² Reza Afrouzian³

1- Department of Biomedical Engineering, Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran Rahmani@tabrizu.ac.ir 2- Department of Biomedical Engineering, Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran 3- Department of Biomedical Engineering, Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran

Keywords :

Acute leukemia.،ALL;،AML;،graph convolutional networks;،microscopic smear analysis;،robust classification

Abstract :

Early triage of acute leukemia remains challenging due to subtle morphologic differences between lymphoid and myeloid blasts and the time-consuming nature of manual review. We present an end-to-end pipeline that integrates graph construction with a tailored deep convolutional architecture for automatic differentiation between acute lymphoblastic leukemia (ALL) and acute myeloid leukemia (AML). The dataset comprises smear images collected from 44 patients; images are resized and normalized, class imbalance is mitigated via GAN-based augmentation, and superpixel-level regions are used to build an adjacency graph whose node features summarize local intensities. A six-layer graph convolutional backbone with batch normalization, dropout, and a terminal softmax performs binary classification. Under a 70/20/10 split with 5-fold cross-validation, the model achieves strong and consistent performance (Accuracy 99.4%, Specificity 97.3%, Kappa 0.85), and remains robust when synthetic white noise is added (accuracy >90% at SNR = 0 dB). Comparative analyses against standard CNN/ResNet/VGG baselines indicate superior accuracy and stability, supporting the efficiency of graph- enhanced representations for this task. These results suggest a practical tool to support pathologists in rapid screening and referral. Future work will extend the framework to multi-class settings (including CML/CLL) and explore alternative augmentation strategies beyond GANs.