Skip to Main Content
Everyone (Public)

BIDS Rising Star/Research In Progress Seminar

Equivariant Imaging Features

You are invited to attend the next seminar in our Research-In-Progress/Rising Star seminar series, hosted by the Section of Biomedical Informatics & Data Science. The seminar will feature John Onofrey, PhD, Assistant Professor of Radiology & Biomedical Imaging and of Urology.


Convolutional neural networks (CNNs) are an essential tool for computer vision and image analysis tasks. However, CNNs are limited by a lack of rotational and flipping equivariance. This deficiency can lead to performance degradation and poor generalization, particularly when dealing with data that lacks explicit orientation, such as in remote sensing and biomedical images. Existing CNN approaches either rely on data augmentation or incorporate modules that extract orientation-sensitive information. However, these methods can significantly escalate the learning cost and may not effectively capture rotation and symmetrical flips. To overcome these challenges, we propose a novel and efficient implementation of a symmetric rotation-equivariant convolution kernel, designed to learn rotation-invariant features while simultaneously compressing the model size. This equivariant kernel can easily be incorporated into any CNN backbone. We validate the ability of our approach to capture equivariance to rotation using the public biomedical MedMNISTv2, EMBED, and NCT-CRC datasets (18 total datasets). Our approach demonstrates improved rotated image classification performance accuracy on all 18 test datasets in both 2D and 3D images, all while increasing efficiency with fewer parameters and reduced memory footprint. Additionally, we evaluate its efficacy on three real-world datasets from remote sensing (NWPU-10 and MTARSI) and natural images (CIFAR-10). Learning equivariant imaging features represents a robust and efficient approach to implementing generalizable CNNs.




Host Organization




Lectures and Seminars


May 20249Thursday