Modelling non-convex fractional order differential equation using dual supervision neural network for medical image quality enhancement

Abstract

Medical imaging is essential for modern cancer screening and diagnosis, but image quality is usually compromised in an attempt to lower patient risk. Traditional computer-aided diagnosis (CAD) systems that employ anomaly detection techniques have increased diagnostic accuracy by assisting radiologists in interpreting medical images. However, existing problems like inconsistent pixel values, computational inefficiencies, and limited generalizability hinder the successful application of AI-based models in real-time clinical settings. Due to distinct pixel intensity variation, medical images necessitate specific transformation; however, model development is complicated by the lack of standard parameter guidelines. Current models' high memory and processing requirements restrict their applicability in settings with limited resources, especially in rural areas. The need for a solution that ensures both diagnostic accuracy and computational efficiency is further highlighted by the fact that noise and artifacts in low-resolution images make it more difficult to diagnose diseases accurately. This study presents a method to improving the quality of medical images by using a non-convex fractional differential equation (NC-FODE). Pixel strength is efficiently calculated by NC-FODE to increase intensity and improve diagnostic relevance. To ensure precise and adaptable parameter setting for different image modalities, a dual supervised neural network (DSNN) is utilized to approximate partial derivatives and set upper bounds for model parameters. Using publicly available radiography datasets, it has been demonstrated that the proposed method greatly enhances image quality across a range of imaging modalities without requiring extensive pre-processing. Real-time processing appropriate for hectic clinical settings is made possible by experimental results showing improved pixel density, decreased noise, and superior computational efficiency. It can be customized for variety of clinical applications since it ensures consistency and reproducibility across different imaging datasets.