Detection of unwarranted CT radiation exposure from patient and imaging protocol meta-data using regularized regression

Ruidi Chen, Ioannis Ch. Paschalidis, Hiroto Hatabu, Vladimir I. Valtchinov, Jenifer SiegelmanEuropean Journal of Radiology Open 2019 · OpenReview

TL;DR

This paper introduces a novel, automated method to detect unwarranted CT radiation exposure by building a predictive model of "normal" radiation dose based on patient and imaging parameters. By identifying scans where the actual dose significantly deviates from the predicted dose, this approach offers a more robust and interpretable way to flag potential overexposure compared to traditional methods, aiding radiation safety committees.

What Problem Does This Paper Address?

Ensuring patient safety in medical imaging, particularly with CT scans, involves monitoring radiation exposure. While CT scans inherently involve radiation, variability in dose can be appropriate due to patient characteristics (like body size) or exam complexity. However, instances of unwarranted high radiation exposure can occur due to protocol errors, equipment malfunctions, or incorrect parameter settings. Traditional methods for identifying these "outliers" often rely on simple statistical thresholds (e.g., dose exceeding the average by a certain multiple of the standard deviation). The problem with these methods is that if the training data itself contains outliers, these can skew the average, masking future outliers. Furthermore, manual review of all CT scans is impractical. This paper addresses the need for an automated, robust, and interpretable method to flag CT exams with potentially unwarranted radiation exposure.

Methods

The core of the proposed method involves a two-stage statistical approach:

  1. Variable Selection with LASSO: To handle the high dimensionality of medical imaging data (many potential predictors), the researchers employed the Least Absolute Shrinkage and Selection Operator (LASSO). LASSO is a penalized regression technique that effectively performs variable selection by shrinking the coefficients of less important predictors to zero, thus identifying a sparse set of key features that are most predictive of CT radiation dose.
  2. Robust Regularized Regression (RR): Instead of standard Ordinary Least Squares (OLS) regression, which is sensitive to outliers, the study used a Regularized Regression (RR) model. RR is designed to be robust to outliers during the training phase. It minimizes the difference between the actual dose and the predicted dose, while also penalizing model complexity. This ensures that the model learns the relationship from the majority of the data, effectively ignoring or downplaying the influence of extreme values.

Outliers are then identified as CT exams where the actual radiation dose deviates significantly (more than 3 standard deviations) from the dose predicted by the trained RR model.

Datasets

The study utilized a retrospective cohort of de-identified CT examination data from a quaternary academic medical center. The initial dataset contained 28 fields for 189,959 CT exams. After pre-processing, the dataset used for analysis comprised 88,566 CT exams with 606 numerically encoded predictors. This dataset is not a newly introduced benchmark but rather a real-world dataset used for training and evaluating the proposed method.

Key Insights and Ideas

  • Outlier Masking: The paper highlights a critical flaw in traditional outlier detection: if the training data contains outliers, these can skew the average and "mask" future outliers, making them appear normal.
  • Robustness is Key: Using a robust regression method (RR) is crucial for building a reliable predictive model of "normal" radiation dose, especially when dealing with potentially contaminated real-world data.
  • LASSO for Interpretability: LASSO not only helps manage high-dimensional data but also provides a more interpretable model by identifying a smaller, more relevant set of predictors.
  • Automated Workflow: The proposed pipeline offers a fully automated solution, reducing the burden on manual review and enabling proactive identification of potential issues.

Key Results

  • The developed method identified 1% of CT exams as outliers.
  • The top predictors for CT radiation dose identified by LASSO included Tube Current, kVp, Weight, Width of collimator, and Reference mAs.
  • Human expert validation of the outlier detection algorithm yielded high accuracy: 91% sensitivity and 85% specificity.
  • The proposed method significantly outperformed alternative methods (a simple "cutoff" method and OLS regression) in terms of the F1 score (0.88 vs. 0.51).
  • The RR-based method detected more outliers, particularly for patients in the "M" and "L" weight categories, which represent the majority of patients.

Limitations and Future Work

  • Data Source Limitations: The study was conducted using data from a single quaternary academic medical center with scanners and software installed before 2012. This limits the immediate generalizability of the findings to other healthcare settings, scanner vendors, and software versions.
  • Future Work: The authors suggest that the methodology can readily incorporate additional variables available in modern CT scanners and Radiation Dose Structured Reports (RDSR). They also propose that the automated outlier detection procedure could be integrated into Radiology Information Systems (RIS) and Picture Archiving and Communication Systems (PACS) to enhance dose monitoring capabilities. Independent testing on data from diverse institutions is also recommended.

Summary updated

BibTeX
@article{chen2019detection,
  author = {Ruidi Chen and Ioannis Ch. Paschalidis and Hiroto Hatabu and Vladimir I. Valtchinov and Jenifer Siegelman},
  title = {{Detection of unwarranted CT radiation exposure from patient and imaging protocol meta-data using regularized regression}},
  journal = {European Journal of Radiology Open},
  year = {2019},
  url = {https://openreview.net/forum?id=3LldN6pYJd},
}
Figure