<p>In computer-aided diagnosis, a key component of the pre-processing phase involves anatomical structure segmentation from medical images. However, it is a difficult work due to poor contrast, non-homogeneous textures, and a lack of labeled data. Anatomical structures in medical images possess well-defined auxiliary information like organ shape and position that may be leveraged to enhance partitioning precision compared to natural images. In this work, we introduced a novel Spatially Regularized Deep Prior (SRDP) network that merges anatomical priors within deep learning models through a dedicated loss function. The SRDP serves as the foundation for the suggested prior loss function. It contains prior organ location and shape data which represent crucial prior data for precise segmentation of organs. Additionally, we propose a dynamic loss function that integrates an earlier and a probability by combining the introduced SRDP loss with traditional probability-based losses. To enhance training, the self-adjusting loss automatically updates the proportion of the probability-based loss to the SRDP loss throughout the learning process. The suggested loss function is versatile and can be used to improve the performance of many different deep segmentation models that are already in use. Using advanced models, such as completely supervised and partially-supervised partitioning frameworks, this work conducted experiments on an open information set in liver segmentation and a confidential information set applied to spleen segmentation confirming the effectiveness of the proposed approach.</p>