DEEP-MISS: Deep Learning-based Fractional Brownian Motion Integrated GAN for Medical Image Segmentation with clinical Sensors

Medical image segmentation (MIS) plays a vital role in different medical applications like analysis, treatment planning, and diagnosis. However, the segmentation accuracy was lower due to the limited availability of labelled medical datasets, mainly for rare ailments. To address this issue, a novel Deep learning-based fractional brownian Motion Integrated Segmentation with clinical Sensors (DEEP-MISS) framework is proposed that integrates fractional Brownian motion (fBm) with the Generative Adversarial Network (GAN) for generating realistic synthetic medical images. The proposed framework initiates with clinical sensor acquired images, which performs reaping and denoising to prepare for the segmentation. The core innovation is the inclusion of fBm integrated GAN architecture (FBI-GAN) for generating the real, texture-rich generated medical images for improving the training dataset. These synthetic samples are then validated to ensure authenticity by comparing them against real images. Moreover, the proposed DEEP-MISS framework employs a Dual attention gates integrated Nested U-Net (DAGIN-Unet) architecture incorporating with channel and position attention gates to focus clinically relevant features. The effectiveness of the proposed DEEP-MISS framework was evaluated with the three datasets such as ISIC2017, LIDC-IDRI and BraTS 2018 based on the segmentation metrics. From the experimental analysis, the proposed DEEP-MISS framework demonstrates improved accuracy in MIS with dice score of 96.13%. The proposed DEEP-MISS framework increases the dice index (DI) by 7.32%, 5.08%, 17.68% and 17.06% for MedSegDiff, UniverSeg, Swin-Unet and MISSFormer respectively.