KARAN: Mitigating Feature Heterogeneity and Noise for Efficient and Accurate Multimodal Medical Image Segmentation

Multimodal medical image segmentation is challenging due to feature heterogeneity across modalities and the presence of modality-specific noise and artifacts. These factors hinder the effective capture and fusion of information, limiting the performance of existing methods. This paper introduces KARAN, a novel end-to-end deep learning model designed to overcome these limitations. KARAN improves feature representation and robustness to intermodal variations through two key innovations: First, KA-MLA, a novel attention block incorporating State Space Model (SSM) and Kolmogorov–Arnold Network (KAN) characteristics into Transformer blocks for efficient, discriminative feature extraction from heterogeneous modalities. Building on KA-MLA, we propose KA-MPE for multi-path parallel feature extraction to avoid multimodal feature entanglement. Second, RanPyramid leverages random convolutions to enhance modality appearance learning, mitigating the impact of noise and artifacts while improving feature fusion. It comprises two components: an Appearance Generator, creating diverse visual appearances, and an Appearance Adjuster, dynamically modulating their weights to optimize model performance. KARAN achieves high segmentation accuracy with lower computational complexity on two publicly available datasets, highlighting its potential to significantly advance medical image analysis.