Abstract

Perfusion cardiovascular magnetic resonance imaging is used to quantify the heart's blood flow, which requires the segmentation of the myocardium, a laborious task. Deep learning-based methods, the most accurate to accomplish this task, still rely on expensive motion correction steps and require large labeled datasets. This paper presents an innovative, efficient approach to myocardial perfusion segmentation, utilizing deep learning techniques without motion correction and with minimal data requirements. Through transfer learning, this methodology leverages the wealth of information available from large, publicly accessible cine magnetic resonance datasets, which provide anatomically analogous images to perfusion ones. This methodology includes normalization and cropping of cine images using a Region-of-Interest detector based on a Markovian, graph-based visual saliency algorithm improved by a sequence of morphological operations. After pretraining a U-net convolutional neural network, a special fine-tuning scheme optimizes its performance. The parameters learned are the starting point for training on a smaller perfusion dataset from the Clinical Hospital of the University of Chile. After an ablation study, the best model is obtained when using both cropping and fine-tuning from the cine dataset, segmenting the left ventricle endocardium with Dice, IoU, and Hausdorff distance of 92.2%, 85.9%, and 5.1 mm respectively, and 95.6%, 91.7%, and 4.6 mm for the left ventricle epicardium. Notably, fine-tuning achieves a Dice of 91.8% for endocardium and 95.2% for epicardium when only 289 perfusion training images are available. These are promising results for developing targeted implementations in real healthcare settings when only small datasets are available.