CMUeXt network model: Analysis and Verification of Insufficient Segmentation Performance on CT and MRI Images

Abstract

Multimodal medical image segmentation is a critical auxiliary technique for clinical diagnosis, but it still confronts notable challenges in addressing noise interference, boundary blurring, and complex anatomical structures in CT, MRI, and other images. This article systematically evaluates the practical performance of the lightweight network CMUeXt, which integrates large convolution kernel and skip fusion module in CT and MRI segmentation tasks. Experiments are conducted on an NVIDIA RTX 4090 GPU using public datasets: CT (lung/ liver), MRI (brain tumor), and mixed-modal datasets. The evaluation adopts Intersection over Union (IoU) and F1- score as core metrics, supplemented by visual validation. Results demonstrate significant limitations of CMUeXt: weak boundary perception ability, ambiguous probability maps, and poor cross-modal generalization. Its large-kernel design and skip fusion module fail to adapt to inherent differences in multimodal data. Future optimizations should integrate dynamic convolution, attention mechanisms, and domain adaptation to enhance its practical clinical applicability.