Simultaneous segmentation of multiple organs from different medical imaging\nmodalities is a crucial task as it can be utilized for computer-aided\ndiagnosis, computer-assisted surgery, and therapy planning. Thanks to the\nrecent advances in deep learning, several deep neural networks for medical\nimage segmentation have been introduced successfully for this purpose. In this\npaper, we focus on learning a deep multi-organ segmentation network that labels\nvoxels. In particular, we examine the critical choice of a loss function in\norder to handle the notorious imbalance problem that plagues both the input and\noutput of a learning model. The input imbalance refers to the class-imbalance\nin the input training samples (i.e., small foreground objects embedded in an\nabundance of background voxels, as well as organs of varying sizes). The output\nimbalance refers to the imbalance between the false positives and false\nnegatives of the inference model. In order to tackle both types of imbalance\nduring training and inference, we introduce a new curriculum learning based\nloss function. Specifically, we leverage Dice similarity coefficient to deter\nmodel parameters from being held at bad local minima and at the same time\ngradually learn better model parameters by penalizing for false\npositives/negatives using a cross entropy term. We evaluated the proposed loss\nfunction on three datasets: whole body positron emission tomography (PET) scans\nwith 5 target organs, magnetic resonance imaging (MRI) prostate scans, and\nultrasound echocardigraphy images with a single target organ i.e., left\nventricular. We show that a simple network architecture with the proposed\nintegrative loss function can outperform state-of-the-art methods and results\nof the competing methods can be improved when our proposed loss is used.\n