Uncertainty Quantification in CNN-Based Surface Prediction Using Shape Priors

Katarína Tóthová, Sarah Parisot, Matthew C. H. Lee, Esther Puyol-Antón, Lisa M. Koch, Andrew P. King, Ender Konukoglu, Marc Pollefeys

Abstract

This paper introduces a probabilistic deep learning framework for medical surface reconstruction that integrates Principal Component Analysis (PCA) shape priors into a Convolutional Neural Network (CNN). Unlike traditional deterministic models that provide a single estimate, this framework quantifies "aleatoric" uncertainty—the error inherent in noisy or blurry medical images. By predicting the parameters of a distribution over PCA weights, the model provides a mean surface along with spatially varying confidence regions. Evaluated on cardiac MRI data from the UK Biobank, the proposed probabilistic model (probPCA) significantly outperforms deterministic baselines in both overlap metrics and point-to-point accuracy.

Introduction

In clinical applications such as surgical planning or disease monitoring, accurate surface reconstruction of organs is critical. However, medical images are frequently affected by low resolution, noise, and artifacts, making boundary identification ambiguous. Most existing deep learning models are deterministic, offering a single "best guess" without indicating where the data might be unreliable. This paper addresses this gap by developing a method to move from simple boundary prediction to a probabilistic representation: "This is the most likely boundary, and here is the range of possible error for every point on that surface."

Methodology

The authors propose a probabilistic approach that combines the feature-learning capabilities of CNNs with the structural constraints of Principal Component Analysis (PCA).

  1. PCA Shape Prior: To ensure anatomically plausible results, the surface is represented as a linear combination of "modes of variation" learned from training data. This prevents the model from generating "jittery" or impossible shapes.
  2. Probabilistic Latent Variables: Instead of predicting PCA weights directly, the CNN outputs the mean $\mu(x)$ and covariance matrix $\Sigma(x)$ of a distribution over these weights.
  3. Architecture and Training: A CNN processes MRI images to estimate these latent distribution parameters. The model is trained to maximize the likelihood of the ground truth surface, using a Kullback-Leibler Divergence (KLD) term to regularize the latent space. During inference, the covariance matrix defines "confidence ellipses" for every vertex on the predicted surface.

Experimental Setup

The study utilizes cardiac MRI volumes from the UK Biobank, focusing on the delineation of myocardium (heart muscle) boundaries.

  • Scale: The model was tested on both Small FOV ($60 \times 60$ pixel crops) and Full FOV ($200 \times 200$ pixel images).
  • Ground Truth: Expert-derived 50-vertex delineations were used as the reference standard.

Key Insights: Spatially Varying Uncertainty

A significant finding of this work is that uncertainty is not uniform across the organ surface. The model's confidence varies based on local image quality and anatomical features.

Figure 6: Zoomed-in views highlighting the relationship between local image features and the magnitude of the predicted uncertainty regions.

The "confidence ellipses" visualized in the results show how the model's doubt changes in both direction and magnitude along the perimeter. This allows clinicians to identify specific regions where the segmentation might require manual adjustment. Furthermore, the authors distinguish between aleatoric uncertainty (data-driven) and epistemic uncertainty (model-driven), focusing on the former to help the model "hesitate" when faced with poor input data.

Results and Discussion

The probabilistic PCA (probPCA) model consistently outperformed deterministic alternatives (detPCA) and direct coordinate regression.

Figure 1: Grid of cardiac MRI slices showing predicted myocardium boundaries (red/cyan) and blue confidence ellipses representing spatially varying uncertainty.
  • Accuracy Boost: On the Full FOV task, probPCA achieved a DICE score of 0.84, a relative improvement of ~7.7% over the deterministic PCA's 0.78.
  • Error Reduction: The Root Mean Square Error (RMSE) for vertex distance was reduced from 4.20mm (deterministic) to 2.59mm (probabilistic).
  • Validation: Qualitative analysis confirmed that ground truth reference points almost always fell within the 99.9% confidence regions generated by the model, validating the statistical grounding of the uncertainty estimates.

Conclusion and Future Work

The integration of PCA priors with probabilistic CNNs provides a robust framework for medical image segmentation with built-in confidence metrics. While currently demonstrated on 2D slices, future work will focus on extending the method to 3D surfaces and exploring non-linear shape models for more complex anatomical structures. Additionally, the authors suggest investigating transfer learning to apply this approach to smaller clinical datasets where labeled data is scarce.

Summary updated

BibTeX
@inproceedings{tothova2018uncertainty,
  author = {Katarína Tóthová and Sarah Parisot and Matthew C. H. Lee and Esther Puyol-Antón and Lisa M. Koch and Andrew P. King and Ender Konukoglu and Marc Pollefeys},
  title = {{Uncertainty Quantification in CNN-Based Surface Prediction Using Shape Priors}},
  booktitle = {ShapeMI @ MICCAI},
  year = {2018},
  url = {https://openreview.net/forum?id=7bbC1nCrUHK},
}
Figure