Efficient convolutional neural networks for acute lymphoblastic leukaemia prediction in computer vision

Area of Science:

• Hematology
• Medical Imaging
• Artificial Intelligence

Background:

• Acute lymphoblastic leukemia (ALL) is a serious hematological malignancy where diagnostic timing impacts survival rates.
• Existing deep learning models for ALL diagnosis face limitations in clinical reliability due to single-model reliance, dataset imbalance, and lack of statistical validation.

Purpose of the Study:

• To propose a robust ensemble framework for improved acute lymphoblastic leukemia diagnosis.
• To enhance the clinical dependability and accuracy of AI-driven medical image analysis for hematological malignancies.

Main Methods:

• An ensemble framework was developed integrating pre-trained Convolutional Neural Networks (CNNs) like DenseNet-121 and ResNet-34 for feature extraction.
• Machine learning classifiers including Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Random Forest (RF), AdaBoost, and Backpropagation Network (BPN) were utilized.
• The framework was evaluated on the C-NMC leukemia dataset comprising 10,661 images.

Main Results:

• The ensemble model achieved 92.5% accuracy and a 93.1% F1-score, surpassing individual CNN models (DenseNet-121 by 5.6%, ResNet-34 by 6.3%).
• The highest Area Under the Curve (AUC) of 0.975 was recorded across classifiers.
• Statistical significance (p < 0.05) was confirmed via t-test and Wilcoxon tests, validating the model's performance improvements.

Conclusions:

• The proposed ensemble framework demonstrates significant potential as an automated clinical decision-support tool for ALL diagnosis.
• This approach reduces manual interpretation errors and expedites diagnosis, enhancing real-world applicability in hematology.
• Combining CNN deep features with ensemble machine learning boosts robustness, sensitivity, and overall performance in clinical workflows.