DEEP LEARNING FOR AUTOMATED DETECTION OF BREAST CANCER IN DEEP ULTRAVIOLET FLUORESCENCE IMAGES WITH DIFFUSION...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep learning models for medical image analysis face challenges due to limited data.
• Diffusion probabilistic models (DPM) can generate high-quality synthetic images from noise.
• Intra-operative margin assessment for breast cancer benefits from accurate image analysis.

Purpose of the Study:

• To investigate the efficacy of DPM for augmenting deep ultraviolet fluorescence (DUV) image datasets.
• To improve breast cancer classification accuracy in intra-operative margin assessment using DUV images.
• To compare DPM-based augmentation with traditional methods like Affine transformations and ProGAN.

Main Methods:

• Applied DPM to generate augmented DUV image data.
• Utilized a pre-trained ResNet for convolutional feature extraction from image patches.
• Employed an XGBoost classifier for patch-level classification and Grad-CAM++ for regional importance mapping.
• Fused patch-level decisions with importance maps for whole-surface prediction.

Main Results:

• Augmenting the DUV dataset with DPM significantly improved breast cancer detection performance.
• Classification accuracy increased from 93% to 97% after DPM augmentation.
• DPM demonstrated superior performance compared to Affine transformations and ProGAN for data augmentation.

Conclusions:

• DPM-based data augmentation is a viable strategy to enhance deep learning performance in medical imaging.
• This approach shows significant promise for improving intra-operative margin assessment in breast cancer surgery.
• The integration of DPM with ResNet and XGBoost offers a powerful tool for accurate cancer detection.