AI-Based Medical Image Segmentation in Oncology

Abstract

Accurate tumour segmentation from medical images is a prerequisite for radiation therapy planning, surgical guidance, treatment response assessment, and radiomics-based prognostication. This study developed and benchmarked five deep learning architectures -- U-Net, Attention U-Net, nnU-Net, Swin-UNETR transformer, and a proposed hybrid CNN-Transformer (CT-Hybrid) -- for multi-organ tumour segmentation across four oncology imaging datasets: brain glioma (BraTS 2021, n = 1,251 MRI), lung tumour (NSCLC-Radiomics, n = 422 CT), liver hepatocellular carcinoma (LiTS, n = 131 CT), and pancreatic ductal adenocarcinoma (MSD-Pancreas, n = 281 CT). The proposed CT-Hybrid architecture-- combining a ResNet-50 convolutional encoder with a Swin Transformer bottleneck and skip-connected decoder - achieved the highest mean Dice similarity coefficient (DSC) across all four datasets (brain 0.892, lung 0.874, liver 0.868, pancreas 0.784), outperforming nnU-Net (the current self-configuring benchmark) by 1.4-3.2% DSC on three of four datasets and matching it on liver. For the clinically critical brain tumour enhancing core sub-region, CT-Hybrid achieved DSC = 0.848 +- 0.042 -- significantly above U-Net (0.782 +- 0.058; p < 0.001) and Attention U-Net (0.814 +- 0.048; p = 0.002). Inference time of the CT-Hybrid (1.84 s per volume on NVIDIA A100) is compatible with real-time clinical workflow integration. These findings establish the CNN-Transformer hybrid paradigm as the new performance standard for multi-organ oncological image segmentation.