PagFormer: Polar Accumulator Grid Integrated into Transformers for Medical Image Segmentation

Hang Zhang, Renjiu Hu, Rongguang Wang, Dongdong Liu, Gaolei Li

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Directed Accumulator, Transformers, Skin Lesion Segmentation, Cardiac Organs

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: Novel encoder-slicer network for medical image segmentation

Abstract:

Recent transformers have made remarkable strides in medical image analysis, enhancing the efficacy of various downstream applications.
Yet, the rich geometric patterns present in medical images offer untapped potential for further refinement.
In this paper, introduce the Polar Accumulator Grid (PAGrid) and seamlessly integrate it into the transformer network, PagFormer, with an aim to improve segmentation performance for elliptical or oval objects in medical images.
Inspired by both the bilateral grid, renowned for its edge-preserving filtering, and the directed accumulator, skilled at integrating geometric shapes into neural networks, PAGrid facilitates geometric-preserving filtering through a symmetric sequence of accumulating, processing, and slicing.
PAGrid preserves elliptical geometry information and promotes the aggregation of global information.
The symmetry between accumulation and slicing in PAGrid allows us to transition from the classic encoder-decoder architecture to an encoder-slicer design, emboddied in the PagFormer.
Additionally, PAGrid's parallelization is managed with CUDA programming, and the back-propagation is enabled for neural network training.
An empirical experiment on three medical image segmentation datasets — specifically, ISIC2017 and ISIC2018 datasets for skin lesions, ACDC datasets for cardiac organs, all of which contains elliptically distributed objects — reveals that our method outperforms other state-of-the-art transformers.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2118