Dynamic kernel generation through hybrid involution and convolution neural networks for leukemia and white blood cell...

Area of Science:

• Hematology
• Computational Biology
• Medical Imaging

Background:

• Accurate blood cancer diagnosis relies on identifying subtle morphological differences in cells, which is challenging with traditional microscopic analysis.
• Automated analysis of microscopic images is crucial for improving the speed and accuracy of leukemia detection and white blood cell (WBC) subtyping.

Purpose of the Study:

• To develop a Hybrid Involutional-Convolutional Neural Network (HICNN) for automated leukemia detection and WBC morphology analysis.
• To enhance diagnostic accuracy by precisely classifying leukemia stages and WBC subtypes using advanced deep learning techniques.

Main Methods:

• Developed a novel HICNN architecture integrating involution layers for adaptive kernel generation and convolutional layers for hierarchical feature extraction.
• Utilized parallel processing within hybrid blocks to improve the discrimination of subtle cellular variations in microscopic images.
• Trained and validated the HICNN on leukemia staging and WBC subtyping datasets.

Main Results:

• Achieved high accuracy: 99.5% for leukemia staging and 98.00% for WBC subtyping, outperforming existing state-of-the-art models.
• Demonstrated model reliability with low Brier scores (0.0019 and 0.0069) and minimal inter-class misclassifications (<2%).
• Observed stable training convergence within 50 epochs, with validation accuracy surpassing 99% by epoch 30.

Conclusions:

• The HICNN effectively addresses challenges in feature discrimination and model calibration for automated hematological diagnostics.
• This framework shows significant promise for reducing diagnostic ambiguity and improving early detection of leukemia and related blood disorders.
• The HICNN represents a significant advancement in AI-driven diagnostic tools for hematopathology.