Optimal Medial Image Segmentation Model Using Neural Fractional Order Differential Equation For Accurate Disease Detection

Accurate image segmentation is an essential process to detect tumors or lesions that make a reliable medical image analysis and diagnosis of diseases. Yet, the tiny size of lesions and high interclass similarity in terms of shape, color, and location make the segmentation a difficult task as traditional multilayer thresholding methods have high computational cost. To address these constraints, this work presents an optimal medical image segmentation model using the neural field ordinary differential equations (NF-ODE). It is used for segmentation to segregate edge representations and UNet-model is used for computational efficient region of interest extraction. To mitigate numerical instability and to resolve the convergence of NF-ODE leaping high-density data, a stochastic conjugate gradient optimization (SCGO) to have training stable toward global optimum. The results on three head-and-neck organs at risk (OAR) datasets show that the proposed NF-ODE+SCGO model achieved better performance than other alternatives, with average DSC of 95.98% (clean) and 96.9% (masked), which is better comparable to UNet, Conv-ODE and NM-ODE counterparts. The average accuracy and DSC of the proposed method is 98.39% and 98.12% on individual datasets, respectively, depicts efficacy in challenging lesion segmentation.