Comparative Study on Attention Unet for Image Segmentation and Propose Inception-Attention based Modification to the Architecture
Joshua Ryan Lawrence Dsouza
Student of MTech DSML, PES University Bengaluru, India
Ruchita Singhania
Mentor,
Great Learning,
Bengaluru, India
Abstract—The advent of Unet in 2015, by Ronneberger et al, the same brought about a uproar in the field of image segmentation and its application in Medical sector. Since then, Numerous efforts have been contributed to by researchers to enhance base architecture and bring about better yield from the algorithm. The main goal for the research being, getting accurate segmentation, yet maintaining the integrity of the result, this has been contributed to by the researchers in the field. This study is undertaken to understand the variations and improvements proposed understand the complexities involved, Implement the base architecture and few of the improvements find the best combination that achieves the desired outcome. A wide known application of Image segmentation algorithms is in Medical Sector. Mainly in identification and isolation of tumors and unknown growths in the human anatomy. The main concern here being, the time constraint, mainly from the first report of the problem to its diagnosis and then treatment. The problem towards which the proposed system adds an edge is to reduce the time needed to check and identify the unknown growth in the regions, using the MR Images of the area Based on the results of the study, the paper discusses about the proposal is to incorporate enhancement, that would yield a better yet accurate segmented result from the base algorithm. From the Research contributed to thus far we have seen attention gates, inception networks with dense inception and other variants, and swin transformer blocks come in as improvements to the base architecture of Unet. Current considerations for the enhancements are Attention Gating mechanism and Inception Mechanism to incorporated into the Base of Unet. The same would be gauged on its performance over the MRI dataset as a Base and later on check its performance other Datasets, the metric for the gauging under consideration would be DICE Scores and or IoU
Index Terms—UNet, Attention Gates, Inception Blocks, Inception Network, Inception V1 Convolution Neural Network, Attention-UNet, Inception UNet.
