How well do U-Net-based segmentation trained on adult cardiac magnetic resonance imaging data generalize to rare congenital heart diseases for surgical planning?

Abstract

Planning the optimal time of intervention for pulmonary valve replacement surgery in patients with the congenital heart disease Tetralogy of Fallot (TOF) is mainly based on ventricular volume and function according to current guidelines. Both of these two biomarkers are most reliably assessed by segmentation of 3D cardiac magnetic resonance (CMR) images. In several grand challenges in the last years, U-Net architectures have shown impressive results on the provided data. However, in clinical practice, data sets are more diverse considering individual pathologies and image properties derived from different scanner properties. Additionally, specific training data for complex rare diseases like TOF is scarce.For this work, 1) we assessed the accuracy gap when using a publicly available labelled data set (the Automatic Cardiac Diagnosis Challenge (ACDC) data set) for training and subsequent applying it to CMR data of TOF patients and vice versa and 2) whether we can achieve similar results when applying the model to a more heterogeneous data base.Multiple deep learning models were trained with four-fold cross validation. Afterwards they were evaluated on the respective unseen CMR images from the other collection. Our results confirm that current deep learning models can achieve excellent results (left ventricle dice of 0.951±0.003/0.941±0.007 train/validation) within a single data collection. But once they are applied to other pathologies, it becomes apparent how much they overfit to the training pathologies (dice score drops between 0.072±0.001 for the left and 0.165±0.001 for the right ventricle).

Unsupervised Domain Adaptation from Axial to Short-Axis Multi-Slice Cardiac MR Images by Incorporating Pretrained Task Networks

“…Thus, the Dice score of the module was 5% better for RV segmentation on the TOF data in comparison to the ACDC data, as the training set included a considerable amount of slices covering this pathology. Simultaneously, the Dice was 6% worse for MYO segmentation of TOF patients compared to the ACDC dataset which aligns with the previous investigations on both datasets from Koehler et al [35].…”

“…can also contribute to the domain gap, which was not extensively investigated in our work so far. Note that in previous work [35], we assessed the generalization gap between pediatric congenital cases and cases from adult cardiology on the same overall data sets.…”

“…4) was chosen based on the original architecture of Ronneberger et al [7]. The hyper-parameters of this network were investigated and optimised with regard to their generalisation capabilities in a previous study [35].…”

“…with w seg = 0.5, which achieved the best performance investigated while working on [35]. To ensure that this weighting is also the best choice for the dataset used in this work, we evaluated the influence of different w seg values on the results within the scope of a sensitivity analysis (cf.…”

Self-supervised motion descriptor for cardiac phase detection in 4D CMR based on discrete vector field estimations

“…Three out of five predicted the ED and ES 'key frame' directly [3,5,13], in contrast to Xue et al [19] and Fiorito et al [7], who classified each frame of the sequence into either diastole or systole. Another obvious approach is to use a LV segmentation approach [12] and infer min/max volume from it as ES and ED.…”