Convolution, Attention and Hybrid Model: A Comparative Study of Deep Models’ Performances on Medical Imaging Segmentation

Abstract

Attention mechanism is recently a trending model in the field of deep learning and Artificial Intelligence. Its success in Natural Language Processing cannot be overemphasized, to the extent that it is being proposed as a replacement for convolution process in computer vision and medical imaging. The idea of replacing convolution with attention mechanism has generated some questions that require prompt attention in the field. In this work we conduct empirical studies by evaluating the performance of both attention-based UNet and convolution-based UNet on the Magnetic Resonant Imaging (MRI) brain tumor dataset. Our findings reveal that the attention UNet model performed better than the pure convolutional UNet. However, the attention model did not totally replace convolution in UNet but compliments it by using its features as input. Furthermore, convolution is still very relevant during upsampling processes, and as skip connector in the decoder phase of attention encoded UNet. Moreover, both mechanisms show superior performance as a hybridized model than their individual state. We deduce that if attention UNet performance is complimented by convolution processes, it suffices to conclude that hybridization of attention and convolution models is the future of medical imaging.