Modernizing Convolutional Segmentation: A ConvNeXt-UNet Architecture for Multi-Modal Medical Image Analysis

Abstract

Manual tumour delineation in radiation oncology is time-intensive and prone to inter-observer variability. While deep learning offers a solution, the widely used U-Net is often limited by 3x3 kernels, which provide an insufficient effective receptive field for diffuse targets. This thesis evaluates an improved ConvNeXt-UNet architecture featuring 7x7 depthwise convolutions, an inverted bottleneck structure, and Layer Normalization to maximize spatial context. Two tasks were studied: Task A (Brain tumour segmentation, Magnetic Resonance Imaging [MRI]) and Task B (Head and neck cancer segmentation, Computed Tomography [CT]). Evaluation was conducted via patient-wise 5-fold cross-validation. The ConvNeXt-UNet significantly outperformed the U-Net in Task A, achieving a mean Dice Similarity Coefficient (DSC) of 0.9328 compared to 0.5496. However, in Task B, the U-Net proved superior (DSC: 0.3000 vs. 0.2558). These findings suggest that architectural advantages may depend on the imaging modality: while large kernels benefit MRI, they may introduce noise in lower-contrast CT scans. The results emphasize the necessity of multi-modal benchmarking to establish the generalizability of medical imaging architectures.