ArachNet: Interpretable Sub-Arachnoid Space Segmentation Using an Additive Convolutional Neural Network

Considering the growing demand for interpretable deep learning models in medical imaging, this paper introduces ArachNet, a novel Convolutional Neural Network (CNN) architecture tailored for spinal sub-arachnoid space segmentation in MR images. Unlike recent relevant approaches, ArachNet is an inherently interpretable model based on the EPU-CNN interpretable framework. The model necessitates an initial image decomposition into perceptually meaningful components representing structural and anatomical features of the spinal region. Each perceptual component is propagated to a dedicated sub-network that outputs a perceptually meaningful 2D representation, enabling the interpretation of the overall inferred segmentation. The experimental evaluation indicates that ArachNet has great application prospects in the navigation of minimally invasive surgical tools, such as vine robot endoscopes, in the spine.