TSDNET: A Tumour Segmentation Network with 3D Direction-Wise Convolution

The segmentation of tumours requires accurate identification and localisation in medical images. With the advent of U-Net and its variants, medical image segmentation has improved by leaps and bounds. Recently, researchers have adapted Transformers for vision tasks by fusing Transformer blocks with the U-Net architecture. Although tumour segmentation results have improved, models have also significantly increased computation costs. To reduce the computation costs while retaining accurate segmentation results, we propose TSDNet, a tumour segmentation network which utilises a novel 3D direction-wise convolution. We conducted experiments on five tumour segmentation datasets and compared the TSDNet model with state-of-the-art tumour segmentation models. Experimental results show that TSDNet performs better at segmenting complex tumour shapes and obtains results at par with ensemble and Transformer models, while also reducing the computation costs.