Medical Image Noise Controller (MINC): Boosting U-NET Based Networks for Superior Segmentation

Segmentation of medical images is very crucial for clinical diagnosis, yet it faces challenges due to factors like noise, complex structures, and blurred boundaries. Rician noise in MRI images makes the problem even more challenging, thus degrading the segmentation accuracy. In this study, Medical Image Noise Controller (MINC) is proposed - a novel architecture to enhance the effectiveness of the U-NET based segmentation networks. The integration of MINC with U-NET, Swin-UNET, and UNETR brings significant improvements in critical segmentation metrics like IoU, Dice coefficient, and MCC. MINC experimental results shows an improvement in IoU from 2.55% to 8.49% on all three networks when tested under noisy conditions. MINC consistently demonstrates its effectiveness in enhancing performance across diverse medical image datasets, even under noisy conditions. The results suggest that MINC is a reliable approach for improving medical image segmentation, with potential for broader application in clinical settings characterized by noisy data.