Rethinking Breast Cancer Diagnosis through Deep Learning Based Image Recognition.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer diagnosis relies heavily on medical imaging interpretation.
• Deep learning offers potential for improving accuracy and efficiency in image analysis.
• Current methods face challenges in achieving maximal diagnostic precision across diverse image modalities.

Purpose of the Study:

• To explore deep learning techniques for breast cancer diagnosis using medical image recognition.
• To investigate image classification and segmentation strategies for optimal accuracy.
• To evaluate the impact of various models, loss functions, and data augmentation on diagnostic performance.

Main Methods:

• Utilized deep learning models including VGGNet19, ResNet50, DenseNet121, EfficientNet v2 for image classification.
• Employed segmentation models such as UNet, ResUNet++, and DeepLab v3.
• Investigated loss functions (binary cross entropy, dice Loss, Tversky loss) and data augmentation techniques.
• Applied methods to X-ray (Mammography), ultrasound, and histopathology images.

Main Results:

• ResNet50 demonstrated superior performance in image classification tasks.
• UNet achieved the best results for image segmentation in both X-ray and ultrasound images.
• Filter-based data augmentation proved effective in enhancing model performance.
• Significant accuracy improvements were observed: 33.3% in X-ray segmentation, 29.9% in ultrasound segmentation, and 22.8% in histopathology classification.

Conclusions:

• Deep learning-based medical image recognition significantly improves breast cancer diagnostic accuracy.
• Specific models like ResNet50 and UNet are highly effective for classification and segmentation, respectively.
• Optimized image recognition strategies enhance diagnostic capabilities across multiple imaging modalities.