Fractal fractional attention UNet based feature enhancement for accurate lung nodule segmentation

Recent deep learning models for medical image segmentation still have challenges and improvements are expected.These include improved multi-scale feature detection, critical region localization, and noise and background reduction. In addition, conventional convolution techniques tend to add computational cost, which can lead to the loss of critical information, especially for low contrast and noisy images, thus affecting the overall segmentation quality.The UNet Fractal Fractional Integral Attention (UNet FFA) mechanism improves feature extraction and noise reduction by combining fractal and fractional integral methods. The fractal identifies structures and patterns at different scales, and fractional integrals of different orders are used to smooth and sharpen the image by the Riemann Liouville derivative. The mask is developed along x and y axes to assign weights for the FFA by using Fractal Fractional Integral (FFI). It highlights important regions with the help of transformation functions.This approach enhances the denoising, detection and segmentation of small lung nodules in their initial stage.The experiments are performed on the LUNA16 and LIDC IDRI datasets.The proposed model performs well with a precision of 0.96 ± 0.01, dice coefficient of 0.98, Intersection over Union (IoU) of 0.97 and recall of 0.95 ± 0.0. It effectively detects lung nodules with diameters ranging from 3mm to over 20 mm. Statistical analysis confirms that the model provides improved dynamic stability and efficiency compared to several recent approaches.The p value also below the significance threshold of 0.05, which means that the results of the model are statistically significant.