Efficient 3D deep learning for myocardial diseases segmentation Khawla Brahim 1,2,3 , Abdul Qayyum 2 , Alain Lalande 2 , Arnaud Boucher 2 , Anis Sakly 3 , and Fabrice Meriaudeau 2 1 National Engineering School of Sousse, University of Sousse, Tunisia 2 ImViA EA 7535 laboratory, University of Burgundy, France {Khawla.Brahim , Abdul.Qayyum , Alain.Lalande , Arnaud.Boucher , Fabrice.Meriaudeau}@u-bourgogne.fr 3 LASEE laboratory, National Engineering School of Monastir, University of Monastir, Tunisia sakly anis@yahoo.fr Abstract. Automated myocardial segmentation from late gadolinium enhancement magnetic resonance images (LGE-MRI) is a critical step in the diagnosis of cardiac pathologies such as ischemia and myocar- dial infarction. This paper proposes a deep learning framework for im- proved myocardial diseases segmentation. In the first step we propose an encoder-decoder segmentation network that generates myocardium and cavity segmentations from the whole volume then followed by a 3D U-Net based on Shape prior identifies myocardial infarction and MVO segmentations from the encoder-decoder prediction. The proposed network, achieves good segmentation performance, as computed by av- erage dice ratio over all predicted substructures, respectively : 'My- ocardium': 96.29%, 'Infarctus': 76.56%, 'No-reflow': 93.12% on our vali- dation EMIDEC dataset consisting of LGE-MRI volumes of 16 patients extracted from the training data. Keywords: LGE-MRI · Myocardial Infarction · Deep Learning. 1 Introduction According to the World Health Organization (WHO) [1], Myocardial Infarction (MI) is one of the main cause of death globally. It essentially develops when oxygen-rich blood flow to the myocardium is suddenly interrupted [2]. However, when revascularization fails, permanent microvascular obstruction phenomenon (MVO, also known as No-reflow) can occur in scar regions. Efficient quantifica- tion of infarcts and MVO is essential for diagnosis and therapy planning. Myocardial Scar Segmentation aims to accurately recognizing myocardial scars areas. Previous prevalent scar segmentation works were often performed using thresholding-based methods, such as the n-standard deviations (n-SD) [3], the full-width at half-maximum (FWHM) [4] and the region growing [5], which are responsive to the regional intensity variation. However, these algorithms fre- quently require a prior knowledge of the expert myocardial location determined