Convolutional Networks Versus Transformers: A Comparison in Prostate Segmentation

Prostate cancer is one of the most common types of cancer that affects men. One way to diagnose and treat it is by manually segmenting the prostate region and analyzing its size or consistency in MRI scans. However, this process requires an experienced radiologist, is time-consuming, and prone to human error. Convolutional Neural Networks (CNNs) have been successful at automating the segmentation of the prostate. In particular, the U-Net architecture has become the de-facto standard given its performance and efficacy. However, CNNs are unable to model long-range dependencies. Transformer networks have emerged as an alternative, obtaining better results than CNNs in image analysis when a large dataset is available for training. In this work, the residual U-Net and the transformer UNETR are compared in the task of prostate segmentation on the ProstateX dataset in terms of segmentation accuracy and computational complexity. Furthermore, to analyze the impact of the size of the dataset, four training datasets are formed with 30, 60, 90, and 120 images. The experiments show that the CNN architecture has a statistical higher performance when the dataset has 90 or 120 images. When the dataset has 60 images, both architectures have a statistical similar performance, while when the dataset has 30 images UNETR performs marginally better. Considering the complexity, the UNETR has 5_ more parameters and takes 5.8_ more FLOPS than the residual U-Net. Therefore, showing that in the case of prostate segmentation CNNs have an overall better performance than Transformer networks.