RCSegNeXt: Efficient Multi-Scale ConvNeXt for Rectal Cancer Segmentation from Sagittal MRI Scans

Abstract

Rectal cancer remains a critical global health challenge, significantly contributing to mor- bidity and mortality worldwide. Magnetic resonance imaging (MRI) in a sagittal plane offers distinct advantages for rectal cancer diagnosis by providing detailed visualization of the rectum and its surrounding anatomy. However, automated segmentation of the rectum and associated tumors remains difficult due to tumor heterogeneity and complex anatom- ical structure, which necessitate multi-scale feature extraction. This study proposes RC- SegNeXt, a novel non-uniform pure-convolutional rectal cancer segmentation architecture that combines shallow anisotropic stages with deep isotropic stages. The anisotropic stages leverage AniNeXt blocks, designed with customized convolutional kernels and pooling op- erations to address the uneven spatial resolution inherent in MRI data. In the isotropic stages, an IsoNeXt block with a Scale-Aware Integration Module (SAIM) enables efficient multi-scale feature fusion by directing information flow through constrained pathways. This design enhances computational efficiency while achieving superior segmentation accuracy. Experiments on two in-house datasets demonstrate the proposed method’s state-of-the-art performances. Code will be open upon acceptance.