An Explainable Deep Learning-Based Feature Fusion Model for Acute Lymphoblastic Leukemia Diagnosis and Severity Assessment

Hajra Khan¹Muhammad Zaheer Sajid²Nauman Ali Khan³Muhammad Fareed Hamid Fareed Hamid⁴Ayman Youssef⁵Areesha Rehman^6*Nour Aburaed⁷

• ¹Department of Computer Software Engineering, Military College of Signals, National University of Science and Technology, Islamabad 44000, Pakistan;, Islamabad, Pakistan
• ²University of Missouri Information Technology Department, Columbia, United States
• ³Department of Computer Software Engineering, Military College of Signals, National University of Science and Technology,, Islamabad, Pakistan
• ⁴Department of Computer Software Engineering, Military College of Signals (MCS), National University of Science and Technology,, Islamabad, Pakistan
• ⁵Department of Computers and Systems, Electronics Research Institute, Cairo, Egypt
• ⁶Bahauddin Zakariya University Faculty of Pharmacy, Multan, Pakistan
• ⁷University of Dubai, College of Engineering and IT,, Dubai, United Arab Emirates

Abstract: Acute Lymphoblastic Leukemia (ALL) is a malignant blood disorder that primarily affects white blood cells, particularly in children, where early and accurate diagnosis is critical for effective treatment and recovery. This study introduces a novel deep learning framework, termed XIncept-ALL, for automated detection and classification of ALL severity levels. The model integrates pre-trained InceptionV3 and Xception networks through feature fusion blocks, enabling robust representation learning. To enhance performance, data auto-augmentation techniques were applied to address class imbalance and reduce overfitting. Furthermore, Grad-CAM visualizations were employed to highlight discriminative regions of ALL cell images, providing interpretability of the model's predictions and validating that the network focuses on clinically relevant features. The proposed system was evaluated using a newly developed private dataset, Pak-ALL, collected from Pakistani hospitals, along with additional datasets from reliable web sources. Extracted features were classified using an XGBoost classifier into four categories: Benign, Early, Pre, and Pro. Extensive experiments demonstrated the superior performance of the proposed framework, achieving an average accuracy of 99.5% on a challenging external dataset (Iranian dataset). The results highlight the efficiency, scalability, and practicality of the XIncept-ALL model for medical image analysis. By offering both improved classification accuracy and interpretable decision support via Grad-CAM, the proposed approach represents a significant advancement toward reliable computer-aided diagnostic systems for ALL, with strong potential to support clinical decision-making and improve patient outcomes.