Improving Trustworthiness of AI Disease Severity Rating in Medical Imaging with Ordinal Conformal Prediction Sets

Charles Lu, Anastasios N. Angelopoulos, Stuart R. PomerantzMICCAI 2022 · DOI · OpenReview

Sure, here's a blog-style summary of the paper "Improving Trustworthiness of AI Disease Severity Rating in Medical Imaging with Ordinal Conformal Prediction Sets":

TL;DR

This paper introduces a novel method called Ordinal Adaptive Prediction Sets (Ordinal APS) that provides statistically rigorous uncertainty guarantees for AI models predicting disease severity in medical images. By leveraging conformal prediction, Ordinal APS generates prediction sets that are guaranteed to contain the true severity with a user-defined probability, enhancing trust and enabling clinicians to identify potentially unreliable AI outputs.

What Problem Does This Paper Address?

The widespread adoption of AI in medical imaging is hindered by a lack of trust, largely due to the perception that deep learning models can fail unpredictably and with misplaced confidence. A core issue is the absence of statistically sound uncertainty quantification. While AI models can achieve high accuracy, clinicians need assurance about the reliability of these predictions, especially in critical decision-making scenarios. This paper addresses the need for robust uncertainty quantification in ordinal classification tasks, specifically for grading disease severity in medical images, where the output categories have a natural order.

Methods

The paper's core contribution is the development and application of Ordinal Adaptive Prediction Sets (Ordinal APS). This method builds upon conformal prediction, a distribution-free technique that provides formal reliability guarantees for predictions. Unlike standard conformal prediction, Ordinal APS is specifically designed for ordinal labels, ensuring that the generated prediction sets are contiguous (e.g., if "moderate" and "severe" are predicted, it won't exclude "mild").

The authors adapt existing conformal prediction algorithms to create prediction sets that are guaranteed to contain the true disease severity with a user-specified probability (1 - α). They propose a novel score function tailored for ordinal labels and an approximate greedy algorithm for computing these prediction sets efficiently. This approach aims to provide clinicians with a measure of confidence in the AI's prediction, allowing them to identify cases where the AI is less certain.

Datasets

The study utilizes a dataset of 409 lumbar spine MRI exams. For each exam, the associated radiology reports were parsed to extract ground truth labels for stenosis severity at 6 vertebral disc levels. This resulted in a total of 6,093 gradings on a four-point ordinal scale: 0 (no stenosis), 1 (mild stenosis), 2 (moderate stenosis), and 3 (severe stenosis).

Key Insights and Ideas

  • Bridging the Trust Gap: The paper demonstrates that statistically rigorous uncertainty quantification, like that provided by Ordinal APS, is crucial for building clinician trust in AI for medical diagnosis.
  • Ordinality Matters: Standard prediction sets can be inappropriate for ordinal tasks. Ordinal APS ensures that predicted sets respect the inherent order of disease severity, leading to more clinically meaningful outputs.
  • High Uncertainty Flags Anomalies: The study reveals a significant correlation between high prediction uncertainty (indicated by larger prediction sets) and the presence of imaging anomalies such as motion artifacts, metallic hardware, and tumors. This suggests that high uncertainty can serve as a valuable signal for potential issues that might degrade AI performance.
  • Practical Clinical Integration: The authors propose a practical use case: flagging high-uncertainty predictions for further review. This can help clinicians focus their attention on more challenging cases and avoid being misled by potentially unreliable AI outputs.

Key Results

  • Empirical Coverage: All three conformal methods evaluated (Ordinal APS, Naive LAC, and Ordinal CDF) empirically achieved the desired marginal coverage guarantees.
  • Set Size Efficiency: Ordinal APS and Naive LAC achieved proper coverage with significantly smaller prediction set sizes compared to Ordinal CDF, making them more practical.
  • Correlation with Anomalies: A manual review of high-uncertainty cases found a significantly higher prevalence of imaging anomalies (17 out of 70 cases) compared to a random sample (5 out of 70 cases). This difference was statistically significant (p < 0.05).
  • Disc-Level Performance: Coverage was found to be inversely correlated with the prevalence of severe stenosis at specific disc levels, with lower lumbar discs showing higher prevalence and slightly lower coverage.

Limitations and Future Work

  • Marginal vs. Conditional Coverage: While the paper guarantees marginal coverage (on average over the population), it acknowledges that conditional coverage (for every individual case) is generally impossible.
  • Clinical Validation: The authors emphasize that prospective evaluation of Ordinal APS in more clinically realistic settings is needed to fully validate its general feasibility and utility.
  • Generalizability: While the study focuses on spinal stenosis, future work could explore the application of Ordinal APS to other ordinal classification problems in medical imaging and beyond.
  • Algorithm Differences: The paper notes that Ordinal APS and Naive LAC produced very similar sets in their experiments due to the high accuracy of the underlying model and a small number of categories. However, they anticipate that these methods would differ more significantly in ordinal problems with more categories or lower model accuracy.

Summary updated

Abstract

The regulatory approval and broad clinical deployment of medical AI have been hampered by the perception that deep learning models fail in unpredictable and possibly catastrophic ways. A lack of statistically rigorous uncertainty quantification is a significant factor undermining trust in AI results. Recent developments in distribution-free uncertainty quantification present practical solutions for these issues by providing reliability guarantees for black-box models on arbitrary data distributions as formally valid finite-sample prediction intervals. Our work applies these new uncertainty quantification methods -- specifically conformal prediction -- to a deep-learning model for grading the severity of spinal stenosis in lumbar spine MRI. We demonstrate a technique for forming ordinal prediction sets that are guaranteed to contain the correct stenosis severity within a user-defined probability (confidence interval). On a dataset of 409 MRI exams processed by the deep-learning model, the conformal method provides tight coverage with small prediction set sizes. Furthermore, we explore the potential clinical applicability of flagging cases with high uncertainty predictions (large prediction sets) by quantifying an increase in the prevalence of significant imaging abnormalities (e.g. motion artifacts, metallic artifacts, and tumors) that could degrade confidence in predictive performance when compared to a random sample of cases.
BibTeX
@inproceedings{lu2022improving,
  author = {Charles Lu and Anastasios N. Angelopoulos and Stuart R. Pomerantz},
  title = {{Improving Trustworthiness of AI Disease Severity Rating in Medical Imaging with Ordinal Conformal Prediction Sets}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2022},
  doi = {10.1007/978-3-031-16452-1\_52},
}
Figure