SNGP Beats MC Dropout With 7x Faster Uncertainty Detection in Pathology
Meleti and Nirschl show that a single-pass Gaussian process head can flag out-of-distribution pathology images without slowing down inference.
ProTakeaways
- Meleti and Nirschl adapt SNGP for uncertainty-aware biomedical image classification.
- Two changes: spectral norm on hidden layers plus a Gaussian Process head with Random Fourier Features.
- Evaluated on six datasets across white blood cells, amyloid plaques, and colorectal histopathology.
- OOD-AUROC reached 0.97 to 1.00 versus 0.39 to 0.84 for deterministic and MC dropout baselines.
- Latency of 0.213 ms per image, roughly 7x faster than MC dropout with 10 passes.
- Open-source PyTorch framework released for reproducible uncertainty-aware digital pathology.
Deep learning models used in digital pathology have a well-known trust problem: they hand out high confidence scores on tissue slides they have never seen before. A new paper by Uma Meleti and Jeffrey J. Nirschl tackles this by adapting Spectral-normalized Neural Gaussian Processes (SNGP) to biomedical image classification, and shipping an open-source PyTorch implementation alongside the results.
The core claim is practical. With two lightweight architectural changes to a standard convolutional network, you get calibrated uncertainty and near-perfect out-of-distribution (OOD) detection, without paying for multiple forward passes at inference time.
The overconfidence problem in pathology
Current deep learning models for digital pathology are often poorly calibrated in out-of-distribution settings, which limits trust and clinical adoption in safety-critical workflows. A classifier trained on blood smears from one lab will still emit a confident label when handed a cardiac slide, or a sample stained on a different scanner.
The usual fixes are Bayesian neural networks, Monte Carlo (MC) dropout, or deep ensembles. They work, but they require multiple stochastic forward passes per image, which is a non-starter for whole-slide images that can contain thousands of tiles. The authors position SNGP as the single-pass alternative that finally makes uncertainty-aware pathology tractable.