Patch-Based Analysis

The full-resolution mammogram is systematically divided into a grid of patches (e.g., 256x256 or 512x512 pixels). A critical parameter is the overlap stride, which ensures that lesions bisected by a tile boundary are fully captured in at least one patch. Typical overlap ranges from 10% to 50%, balancing computational load against detection sensitivity for small objects like microcalcification clusters.

Downsampling a 4000x5000 pixel mammogram to fit a standard network input destroys subtle indicators of malignancy. Patch-based analysis preserves native spatial resolution, allowing a deep convolutional neural network to learn discriminative features such as:

• Fine spiculation radiating from mass margins
• Individual microcalcification morphology and pleomorphism
• Subtle architectural distortion patterns in dense tissue

Individual patch predictions are not the final output. A global aggregation module synthesizes patch-level features into a whole-image decision. Techniques include:

• Multiple Instance Learning (MIL): Treats the image as a bag of patches, learning which patches are diagnostically relevant
• Spatial pooling: Combines feature vectors from all patches before a final classifier
• Recurrent aggregation: Sequentially processes patches to model spatial dependencies across the breast

In a typical mammogram, over 99% of extracted patches represent normal tissue. This extreme class imbalance can cause models to predict 'normal' universally. Mitigation strategies include:

• Hard negative mining: Actively selecting patches the model finds difficult to classify correctly
• Focal loss: A loss function that down-weights easy negative examples, forcing the model to focus on rare positive patches containing lesions
• Patch-level data augmentation: Oversampling minority class patches during training

Processing thousands of overlapping patches per image is computationally intensive. Optimizations include:

• Fully convolutional networks: Replace dense layers to process arbitrary patch sizes efficiently
• Tissue masking: Pre-segment the breast region to skip background patches entirely
• Cascaded architectures: A fast, low-sensitivity model screens all patches; only suspicious candidates pass to a slower, high-accuracy model These techniques are essential for meeting clinical worklist prioritization latency requirements.

A single patch size cannot capture both fine calcifications and large masses. Multi-resolution analysis extracts patches at multiple scales:

• High-resolution patches (small field of view): Capture microcalcification morphology
• Low-resolution patches (large field of view): Capture mass margins and surrounding architectural distortion Feature vectors from each scale are concatenated or fused via attention mechanisms, providing the model with both local texture and regional context.