Slide-Level Classification

Slide-level classification operates under weak supervision because the only label available is for the entire gigapixel image, not for individual regions. The model must learn to identify diagnostically relevant morphology without pixel-level annotations.

• Multiple Instance Learning (MIL) is the dominant framework
• The slide is a "bag" containing thousands of unlabeled patches
• A positive label indicates at least one region contains disease
• Eliminates the bottleneck of exhaustive manual annotation

Modern architectures use a trainable attention mechanism to learn which patches contribute most to the final diagnosis. The network assigns a weight to each patch, creating an interpretable heatmap of diagnostically salient regions.

• Attention pooling replaces naive averaging or max pooling
• Enables the model to ignore irrelevant stroma and glass background
• Produces a weighted sum of patch-level feature vectors
• Weights can be visualized as a probability heatmap for interpretability

A pre-trained convolutional neural network or vision transformer first encodes every extracted patch into a compact feature vector. This step reduces each image tile to a dense numerical representation before aggregation.

• Common backbones: ResNet-50, ViT-B/16, pathology-specific foundation models
• Features are typically extracted from the penultimate layer
• Self-supervised pre-training on histology data significantly boosts performance
• Feature extraction is the most computationally intensive step

The final output is a single diagnostic probability mapped to clinically meaningful categories. This binary or multi-class prediction directly supports pathologist workflow by triaging cases or providing a second read.

• Binary: Cancerous vs. Benign
• Multi-class: Tumor subtyping (e.g., adenocarcinoma vs. squamous cell carcinoma)
• Grading: Low-grade vs. High-grade based on morphological patterns
• Outputs integrate into Laboratory Information Systems (LIS) via standard APIs

Processing a whole slide image requires a tiled approach due to memory constraints. The image is decomposed into thousands of overlapping or non-overlapping patches, each processed independently before global aggregation.

• Patch extraction at multiple magnifications (e.g., 20x, 40x)
• Tissue masking to discard empty glass background
• Stain normalization to reduce inter-lab color variability
• Parallelized GPU inference across patch batches for throughput

Performance is measured using metrics that account for the clinical consequences of false negatives and false positives. Receiver Operating Characteristic (ROC) analysis is standard, but task-specific metrics are critical.

• Area Under the ROC Curve (AUC) for overall discriminative ability
• Sensitivity at high specificity (e.g., 95%+) to minimize false alarms
• F1 Score for imbalanced datasets where disease is rare
• Slide-level cross-validation prevents patient-level data leakage