Published August 9, 2025 | Version v1

Equivariant Spatiotemporal Transformers for Enhanced Characterization of Malignancies in Dynamic PET

Authors/Creators

Description

This document presents a research proposal titled "E(3)-Equivariant Spatiotemporal Transformers with Shannon Entropy for Robust Malignancy Detection in Dynamic PET Imaging." It introduces the E(3)-Equivariant Spatiotemporal Transformer with Entropy (EST-E), a novel deep learning framework designed to enhance malignancy detection in dynamic Positron Emission Tomography (PET) imaging. The methodology integrates E(3)-equivariant neural networks for robustness to rotations and translations, a transformer architecture for joint spatiotemporal feature modeling, and multi-scale Shannon entropy to capture tissue heterogeneity indicative of malignancy. The proposal details an advanced preprocessing pipeline, including motion correction, SUV normalization, and data augmentation, alongside a 12-layer hybrid transformer and post-processing for uncertainty estimation and interpretability. The document is formatted in IEEEtran two-column conference style, featuring a compacted algorithm, a single-column graph comparing AUC performance, and a three-color gradient visualization of entropy. Appendices provide rigorous mathematical formulations for equivariance, entropy, and transformer mechanisms. This work aims to improve automated malignancy detection in 4D medical imaging, with potential applications to dynamic MRI and CT. The proposal is intended for researchers and practitioners in medical imaging, deep learning, and radiomics.

Files

Equivariant_Spatiotemporal_Transformers_for_Enhanced_Characterization_of_Malignancies_in_Dynamic_PET.pdf

Additional details

Dates

Available
2025-08-09