Explainable medical image clustering

Zhi Xu, Qi Zhao, Xueqi Liu, Kirby Madden-Hennessey, Jordan Urbani, Yun Fu

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

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: Explainable AI, unsupervised learning

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

Abstract: Image classification stands as a pivotal task within the realm of computer vision, entailing the assignment of labels to entire images. Nonetheless, the complete supervision necessary for such microwell image classification demands extensive annotations, a process that can prove time-intensive to accomplish. Furthermore, situations arise where delving into the intrinsic attributes of data is desired, even when data labels remain uncertain. In this paper, we introduce a cell dataset that captures the developmental trend of cancer cells, along with T Cells, under the influence of diverse experimental conditions medications. Concurrently, we present an approach to both cluster input images and elucidate the rationale behind their grouping. To achieve this, we leverage a U-net encoder for individual microwell image information encoding and a multi-head attention layer for information encapsulation across different time points. Subsequent to clustering, we employ various techniques to expound upon our clustering outcomes. Specifically, we utilize Grad-CAM for visual explication, coupled with human-friendly textual generation aimed at facilitating comprehension of trends within each cluster. Our study encompasses a comparison of diverse architectural models on our proposed dataset, conclusively demonstrating the superior performance of its architecture. Experimental analyses and ablation studies further substantiate the advantages conferred by our innovative architectural approach.

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: 3713