Nnet: N-type neural network for accurate segmentation of brain structures in MRI

Abstract

Accurate segmentation of brain structures in magnetic resonance imaging plays a key role in the diagnosis and treatment of idiopathic Normal Pressure Hydrocephalus (iNPH). However, existing spatial domain-based methods are limited by the complexity of brain anatomical structures and the low contrast of medical images, making it difficult to achieve high-precision tissue boundary segmentation. To address this challenge, this paper re-examines the problem of brain structure segmentation from a new perspective in the frequency domain and proposes an N-type neural network architecture (Nnet). Nnet efficiently extracts low-frequency content information and high-frequency edge texture information in images through the coordinated operation of three parallel branches, thereby achieving accurate positioning of the target boundary. In addition, Nnet integrates the frequency domain online enhancement (FOE) module and the feature communication (FCM) module to further optimize the segmentation performance. The FOE module regulates the competition and cooperation between channels through the gating mechanism, effectively constructs the relationship between frequency domain features and reduces local information distortion. The FCM module uses the “growth” and “communication” mechanisms to fundamentally improve feature consistency by placing high-level and low-level features at the same level for communication, eliminating the semantic gap problem between branches. The results of this paper on two public brain MRI T1 sequence datasets (MALC dataset and IBSR dataset) show that the Dice coefficient of Nnet is improved by 1.29% and 1.47% on average, respectively, which is significantly better than the current advanced methods.