Blast cell segmentation and leukemia classification using hybrid Deep Kronecker WideResNet using blood smear images

Area of Science:

• Medical Imaging and Artificial Intelligence
• Hematology and Oncology

Background:

• Acute Lymphoblastic Leukemia (ALL) is a critical condition affecting bone marrow and white blood cells.
• Early ALL detection is vital for effective patient treatment and management.
• Classifying ALL from Peripheral Blood Smear (PBS) images is challenging due to cell size and shape variability.

Purpose of the Study:

• To develop an advanced computational model for accurate ALL classification from PBS images.
• To address the difficulties in segmenting and classifying leukemia cells in microscopic images.

Main Methods:

• Preprocessing of blood smear images using adaptive Gaussian filtering to reduce noise.
• Segmentation of blast cells using RefineNet, followed by image augmentation (flipping, resizing, rotation).
• Feature extraction including Binary Pattern of Phase Congruency (BPPC) with Discrete Cosine Transform (DCT) and statistical features.
• Classification of leukemia subtypes (early Pre-B, Pre-B, Pro-B, Hematogones) using the proposed Deep Kronecker Wide Residual Network (DKWRN).

Main Results:

• The DKWRN model achieved high performance metrics: 92.12% accuracy, 91.40% True Negative Rate (TNR), 90.36% recall, 91.56% precision, and 90.96% F1-score.
• DKWRN demonstrated significant improvements over existing methods, including Bayesian-based optimized CNN (BO-ALLCNN), SVM, ResNet-18, ResNet-152, DKN, and WRN.
• The model's accuracy surpassed BO-ALLCNN by 5.91% and SVM by 4.63%.

Conclusions:

• The proposed DKWRN model offers a robust and accurate solution for classifying leukemia from PBS images.
• This approach enhances the potential for earlier and more precise diagnosis of ALL.
• The DKWRN's superior performance highlights its promise in clinical applications for hematological disorder detection.