An Intuitionistic Fuzzy Logic Driven UNet architecture: Application to Brain Image segmentation

TL;DR

This work tackles brain MRI segmentation under boundary uncertainty caused by partial volume effects by introducing IF-UNet, a UNet variant that processes inputs in intuitionistic fuzzy form using membership, non-membership, and hesitation degrees. By fuzzifying image data as $x_j^{IFS}=(\mu_B(x_j),\nu_B(x_j),\pi_B(x_j))$ and applying Sugeno negation with a tunable parameter $\lambda$, the encoder captures uncertainty while the decoder maintains spatial localization. The authors demonstrate, on the IBSR dataset, that IF-UNet consistently outperforms baseline UNet and Attention UNet in terms of accuracy $AC$, Dice $DC$, and intersection over union $IoU$, with best results at $\lambda=1.2$ (e.g., $AC=0.9924$, $DC=0.9892$, $IoU=0.9788$). While the approach increases model complexity and inference time due to additional fuzzy components, the improved boundary delineation and robustness to uncertainty have meaningful implications for clinically reliable brain tissue segmentation.

Abstract

Accurate segmentation of MRI brain images is essential for image analysis, diagnosis of neuro-logical disorders and medical image computing. In the deep learning approach, the convolutional neural networks (CNNs), especially UNet, are widely applied in medical image segmentation. However, it is difficult to deal with uncertainty due to the partial volume effect in brain images. To overcome this limitation, we propose an enhanced framework, named UNet with intuitionistic fuzzy logic (IF-UNet), which incorporates intuitionistic fuzzy logic into UNet. The model processes input data in terms of membership, nonmembership, and hesitation degrees, allowing it to better address tissue ambiguity resulting from partial volume effects and boundary uncertainties. The proposed architecture is evaluated on the Internet Brain Segmentation Repository (IBSR) dataset, and its performance is computed using accuracy, Dice coefficient, and intersection over union (IoU). Experimental results confirm that IF-UNet improves segmentation quality with handling uncertainty in brain images.

Figures (6)

• Figure 1: Intuitionistic fuzzy data in form of (a): membership, (b): non-membership and (c): hesitation degree.
• Figure 2: A framework for IF-UNet architecture. It processes the input image in terms of intuitionistic fuzzification, where the image is represented in term of membership, non-membership, and hesitation degree. The encoder part processes the data in terms of membership and non-membership, while the decoder part follows the UNet architecture.
• Figure 3: The training performance of UNet, Attention UNet and IF-UNet during the training process in terms of $AC$,$DC$,and $IoU$ are shown in Fig. \ref{['fig:Fig3']}\ref{['fig:3a']}--\ref{['fig:3c']} respectively. The various value of $\lambda(=0.5,0.9,1.2,1.5)$ in the Sugeno negation function is considered in the study for IF-UNet architectures, and it show better results in comparison to UNet and Attention UNet. It can be shown that the segmentation performance of IF-UNet varies with different $\lambda$ values, yet consistently better performance in comparison to UNet and Attention UNet.
• Figure 4: The comparison of segmentation performance of UNet, Attention UNet and the proposed IF-UNet with Sugeno negation function, measured in terms of $AC, DC$, and $IoU$ on the IBSR dataset using different values of $\lambda$ in the Sugeno negation function. The experimental results show that the proposed IF-UNet consistently achieves improved segmentation accuracy across all tested values of $\lambda$, demonstrating its effectiveness in capturing vague and qualitative information compared to UNet and Attention UNet.
• Figure 5: The comparison of training validation of UNet, Attention UNet and IF-UNet with the Sugeno negation function, measured in terms of $AC\_val,DC\_val$, and $IoU\_val$ on IBSR dataset for various of $\lambda (=0.5,0.9,1.2,1.5)$ in Sugeno negation function. In training validation, the results show that the IF-UNet consistently achieves improved segmentation validation in comparison to UNet and Attention UNet.