The self-configuring nnU-Net has achieved leading performance in a large range of medical image segmentation challenges. It is widely considered as the model of choice and a strong baseline for medical image segmentation. However, despite its extraordinary performance, nnU-Net does not supply a measure of uncertainty to indicate its possible failure. This can be problematic for large-scale image segmentation applications, where data are heterogeneous and nnU-Net may fail without notice. In this work, we introduce a novel method to estimate nnU-Net uncertainty for medical image segmentation. We propose a highly effective scheme for posterior sampling of weight space for Bayesian uncertainty estimation. Different from previous baseline methods such as Monte Carlo Dropout and mean-field Bayesian Neural Networks, our proposed method does not require a variational architecture and keeps the original nnU-Net architecture intact, thereby preserving its excellent performance and ease of use. Additionally, we boost the segmentation performance over the original nnU-Net via marginalizing multi-modal posterior models. We applied our method on the public ACDC and M &M datasets of cardiac MRI and demonstrated improved uncertainty estimation over a range of baseline methods. The proposed method further strengthens nnU-Net for medical image segmentation in terms of both segmentation accuracy and quality control.

Quantitative MRI (qMRI) of the heart has become an important clinical tool for examining myocardial tissue properties. Because heart is a moving object, it is usually imaged with electrocardiogram and respiratory gating during acquisition, to “freeze” its motion. In reality, gating is more-often-than-not imperfect given the heart rate variability and nonideal breath-hold. qMRI of the heart, consequently, is characteristic of varying image contrast as well as residual motion, the latter compromising the quality of quantitative mapping. Motion correction is an important step prior to parametric mapping, however, a long-standing difficulty for registering the dynamic sequence is that the contrast across frames varies wildly: depending on the acquisition scheme some frames can have extremely poor contrast, which fails both traditional optimization-based and modern learning-based registration methods. In this work, we propose a novel framework named DisQ, which Disentangles Quantitative mapping sequences into the latent space of contrast and anatomy, fully unsupervised. The disentangled latent spaces serve for the purpose of generating a series of images with identical contrast, which enables easy and accurate registration of all frames. We applied our DisQ method to the modified Look-Locker inversion recovery (MOLLI) sequence, and demonstrated improved performance of T₁ mapping. In addition, we showed the possibility of generating a dynamic series of baseline images with exactly the same shape, strictly registered and perfectly “frozen". Our proposed DisQ methodology readily extends to other types of cardiac qMRI such as T₂ mapping and perfusion.