Efficient Active Learning for Realtime MRI of Patients with Congenital Heart Disease

Kurzfassung

Here we present our deep learning workflow to reduce the time and effort to annotate and evaluate very large amounts of image data by the use of Active Learning. Especially for rare diseases the availability and time of qualified medical experts is often scarce which is why we focus on minimizing the effort required by those experts while assuring high evaluation quality. State-of-the-art technologies enable medical imaging with ever-increasing spatial and temporal resolution. On the one hand, this allows more precise diagnoses and sharpens our understanding of the human organism. On the other hand, this presents us with the problem of enormous amounts of data that have to be analysed. Since the development of real-time MRI, that is intensively used in cardiac MRI, manual evaluation of the images is no longer feasible. Especially in such a sensitive area as clinical diagnostics, the interaction between humans and software is of enormous importance. For the evaluation of clinically important parameters such as stroke volume or aortic blood flow, the use of semantic segmentation is useful, as it is similar to manual evaluation. Much of the present research in the field of automatic cardiac MRI analysis has focused on healthy hearts and partly on common diseases. However, higher temporal resolutions are particularly interesting to improve the understanding of very severe but rather rare heart diseases. Rare congenital heart disease benefits the least of healthy training data. The small amount of specialised training data greatly challenges automatic evaluation. In this project we focus on evaluating real-time MRI of univentricular heart patients. These patients have only one working ventricle and surgeries lead to a dazzling array of anatomies. In our implemented workflow we start with a relatively small but well-balanced dataset of individual 2D cardiac MRI data, which were annotated by medical experts in the field of univentricular hearts (< 1000 images from < 60 patients). With this small training dataset, we train a UNet using the nnUNet framework. Subsequently, we can evaluate entire real-time MRI datasets frame by frame using this model. We then compute metrics that evaluate the predictive quality of a single image relative to its spatial and temporal neighbours, such as an interslice Dice score. In addition, we can incorporate physiological knowledge into the evaluation of individual predicted segmentations. Using these scores, we can filter particularly poorly predicted segmentations and have them corrected by the medical experts. This allows us to efficiently update the training dataset and re-train the model used in the previous iteration. Since we do not determine the performance of our tool by a limited test data set, as is usually the case, but estimate it on the basis of each newly evaluated patient, we avoid overfitting on the test data and can enable continuous further learning. This is particularly useful in the domain of congenital heart disease, as anatomies of patient’s hearts can vary heavily. To evaluate parameters like the stroke volume or blood flow, we first synchronize the respiratory and cardiac phase by finding slice-wise extrema in the predicted segmentations. First results of these methods show valid results for parameters like respiratory dependent stroke volume of univentricular patients, as well as aortic blood flow.