Spectral-aware CNN with learnable biorthogonal units and depthwise convolutions for multi-class blood cell...

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence

Background:

• Accurate peripheral blood cell classification is crucial for early diagnosis of diseases like leukemia and anemia.
• Existing methods may face challenges in feature retention and computational efficiency.

Purpose of the Study:

• To propose a novel hybrid deep learning model for multi-class blood cell classification.
• To enhance the accuracy and efficiency of blood cell classification for improved disease diagnosis.

Main Methods:

• Developed a Spectral-Aware Convolutional Neural Network (CNN) model incorporating Learnable Spectral Biorthogonal Downsampling Units (LSBDUs).
• Replaced conventional pooling layers with wavelet-inspired LSBDUs for superior feature preservation.
• Integrated Depthwise Separable Convolutions to reduce computational overhead and training costs.
• Utilized a balanced dataset of 17,092 images across eight blood cell classes, employing stratified data splitting, advanced augmentation, and label smoothing.

Main Results:

• Achieved an overall classification accuracy of 99.18% on the blood cell dataset.
• Demonstrated superior class-wise performance compared to existing methods.
• Showcased improved generalization across all classes without overfitting, validating the model's robustness.

Conclusions:

• The proposed Spectral-Aware CNN model with LSBDUs and depthwise separable convolutions offers a highly accurate and efficient solution for multi-class blood cell classification.
• This advancement holds significant potential for improving early disease detection and diagnosis in clinical settings.
• The model's design effectively balances feature preservation with reduced computational complexity.