Uncertainty Estimation for Heatmap-Based Landmark Localization.

Automatic anatomical landmark localization has made great strides by leveraging deep learning methods in recent years. The ability to quantify the uncertainty of these predictions is a vital component needed for these methods to be adopted in clinical settings, where it is imperative that erroneous predictions are caught and corrected. We propose Quantile Binning, a data-driven method to categorize predictions by uncertainty with estimated error bounds. Our framework can be applied to any continuous uncertainty measure, allowing straightforward identification of the best subset of predictions with accompanying estimated error bounds. We facilitate easy comparison between uncertainty measures by constructing two evaluation metrics derived from Quantile Binning. We compare and contrast three epistemic uncertainty measures (two baselines, and a proposed method combining aspects of the two), derived from two heatmap-based landmark localization model paradigms (U-Net and patch-based). We show results across three datasets, including a publicly available Cephalometric dataset. We illustrate how filtering out gross mispredictions caught in our Quantile Bins significantly improves the proportion of predictions under an acceptable error threshold. Finally, we demonstrate that Quantile Binning remains effective on landmarks with high aleatoric uncertainty caused by inherent landmark ambiguity, and offer recommendations on which uncertainty measure to use and how to use it. The code and data are available at https://github.com/schobs/qbin.

Machine learning cardiac-MRI features predict mortality in newly diagnosed pulmonary arterial hypertension.

Aims: Pulmonary arterial hypertension (PAH) is a rare but serious disease associated with high mortality if left untreated. This study aims to assess the prognostic cardiac magnetic resonance (CMR) features in PAH using machine learning. Methods and results: Seven hundred and twenty-three consecutive treatment-naive PAH patients were identified from the ASPIRE registry; 516 were included in the training, and 207 in the validation cohort. A multilinear principal component analysis (MPCA)-based machine learning approach was used to extract mortality and survival features throughout the cardiac cycle. The features were overlaid on the original imaging using thresholding and clustering of high- and low-risk of mortality prediction values. The 1-year mortality rate in the validation cohort was 10%. Univariable Cox regression analysis of the combined short-axis and four-chamber MPCA-based predictions was statistically significant (hazard ratios: 2.1, 95% CI: 1.3, 3.4, c-index = 0.70, P = 0.002). The MPCA features improved the 1-year mortality prediction of REVEAL from c-index = 0.71 to 0.76 (P ≤ 0.001). Abnormalities in the end-systolic interventricular septum and end-diastolic left ventricle indicated the highest risk of mortality. Conclusion: The MPCA-based machine learning is an explainable time-resolved approach that allows visualization of prognostic cardiac features throughout the cardiac cycle at the population level, making this approach transparent and clinically interpretable. In addition, the added prognostic value over the REVEAL risk score and CMR volumetric measurements allows for a more accurate prediction of 1-year mortality risk in PAH.