Abstract
There is a need of accurately segmenting the magnetic resonance (MR) images for diagnosis as well as treatment planning. It is, however, difficult due to low contrast, indistinct borders, intensity inhomogeneity, as well as acquisition noise. Rician noise in MR images reduces accuracy and reliability of segmentation of structures in images that are clinically relevant. Yet it is still challenging to maintain robustness under noisy imaging conditions. To remedy this limitation, we propose an Attention-based Noise-aware Segmentation Network (ANSN) which extends Attention U-Net model by integrating a Noise Suppression Filter (NSF) in the encoder along with attention-guided feature refinement in the decoder. The ACDC, SCD and TCGA-LGG datasets are used to evaluate the proposed framework. With ACDC at 10% Rician noise, ANSN obtains an improvement in the metrics Dice Score and Mean IoU of 6.66% and 5.34% respectively compared to the Attention U-Net. Similarly robust results are observed on the SCD dataset, as ANSN gains 35.68% Dice Score and 15.07% Mean IoU on 10% level of noise. Despite these gains, ANSN requires only 7.40 million parameters, representing an 80.2% reduction compared with Attention U-Net, while maintaining a 50.6 ms inference time on a Tesla T4 GPU.