Does Prior Knowledge Enhance Accuracy of Deep Learning-assisted Semantic Segmentation?

Abstract

In medical image analysis, precise and automated initial segmentation of anatomical organs/structures is vital. This would allow radiologists and physicians to develop a reliable, accurate, cost-effective, and time-effective diagnostic framework. Deep learning techniques have shown promise in automated initial segmentation, but they rely on overall loss function minimization/optimization, which may not result in the production of a realistic/valid shape variation of the target structure. As a result of the lack of explicit form restrictions or priors, deep learning techniques would show significant errors/outliers when it came to the underlying shape/morphology aspects of target structures. The goal of this study was to see how past knowledge was incorporated into deep learning models and how that affected initial segmentation accuracyTo do this, a cutting-edge deep learning model was trained on three independent datasets, including hippocampus segmentation from MR images, kidney segmentation from CT images, and bone/skull segmentation from brain MR images, both with and without prior information. Prior knowledge was added into the deep learning model as an input channel in the form of a solution from an atlas-based technique and a shape model. The performance of the deep learning model without shape prior (DL), the deep learning model with shape prior from an atlas-based solution (DL-Atlas), and the shape model were compared (DL-Shape). The examination also took into account the atlas-based strategy. When prior knowledge in the form of the shape model (DL-Shape) was incorporated into the deep learning model, segmentation accuracy improved to 88.6±1.7% percent (Hippocampus), 91.6±1.5% percent (Kidney), and 86.7±2.5% percent (Bone). These indices improved to 88.6±1.7% percent (Hippocampus), 93.9±1.4% percent (Kidney), and 89.6±2.7% percent (Bone). The statistical study revealed that the DL and DL-Shape models varied significantly. Comparing the performance of the deep learning model with and without prior knowledge demonstrated the significant benefits of incorporating prior shape knowledge into deep learning algorithms for semantic structure segmentation.