Stochastic Dilated Residual Ghost Model for Breast Cancer Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Standardizing breast cancer screening is a significant challenge in modern medicine.
• Deep learning models show promise in distinguishing between benign and malignant tumors.
• Histopathology images present difficulties due to color variations from staining and biopsy materials, leading to inaccurate diagnoses.

Purpose of the Study:

• To address overfitting and color divergence in deep learning models for breast cancer classification.
• To develop a novel method for extracting and enhancing fine-grained features in histopathology images.
• To improve the reliability of deep learning models in recognizing very small objects for accurate breast cancer diagnosis.

Main Methods:

• Strain normalization and adding extra components were used to rectify overfitting and color divergence.
• A multiscale stochastic and dilation unit was developed for extracting fine-grained features at various image scales.
• A stochastic pooling block with convolution and identity mapping was employed to maintain feature mappings and reduce dimensionality.

Main Results:

• The proposed methods successfully addressed overfitting and color divergence issues.
• The multiscale unit effectively extracted edges, contours, and improved color accuracy.
• The enhanced model demonstrated superior performance compared to existing methods, achieving an area under the curve of 96.15%.

Conclusions:

• The developed deep learning approach offers a more accurate and reliable method for breast cancer screening.
• Addressing image variations and enhancing feature extraction are crucial for improving diagnostic performance.
• The proposed techniques show significant potential for standardizing and advancing breast cancer diagnosis through artificial intelligence.