An efficient cell micronucleus classification network based on multi-layer perception attention mechanism

Area of Science:

• Biomedical Imaging
• Computational Toxicology
• Oncology

Background:

• Cellular micronucleus detection is crucial for pathological toxicology and early cancer diagnosis.
• Challenges include tiny targets, high similarity, limited data, and class imbalance in image detection.
• Existing methods struggle with these complexities, necessitating advanced computational approaches.

Purpose of the Study:

• To develop a lightweight yet effective deep learning network for cellular micronucleus image detection.
• To address challenges like small object identification, data scarcity, and imbalanced datasets.
• To improve the accuracy and reliability of micronucleus detection for diagnostic applications.

Main Methods:

• Proposed a lightweight network, MobileViT-MN, integrating a multilayer perceptual attention mechanism.
• Employed data augmentation and transfer learning based on domain adaptation to handle data limitations.
• Introduced a novel Deep Separation-Decentralization module utilizing attention and deep separable convolution.

Main Results:

• MobileViT-MN demonstrated outstanding performance on an augmented dataset.
• Achieved an average accuracy (Avg_Acc) of 0.933, F1 score of 0.971, and ROC score of 0.965.
• Outperformed classical algorithms in classification performance for micronucleus detection.

Conclusions:

• MobileViT-MN effectively addresses the challenges in cellular micronucleus image detection.
• The proposed model offers a significant advancement in accuracy and reliability for toxicological and cancer diagnostic applications.
• The integration of attention mechanisms and novel network modules contributes to superior performance.