E-DDPG: Dual-Objective Deep Deterministic Policy Gradient for MRI Acceleration and Disease Classification

ICLR 2026 Conference Submission17325 Authors

Keywords: Magnetic Resonance Imaging, Reinforcement Learning, Disease Diagnosis, Sampling Strategy

TL;DR: The paper offers a novel mechanisms to the RL in the context of MRI.

Abstract: Long acquisition times remain a major challenge in clinical MRI, where a fundamental trade-off exists between the acceleration achieved through undersampling and the diagnostic utility of the reconstructed images. We cast the problem of acquiring MRI data within a fixed time budget as a discrete reinforcement learning (RL) task and propose an algorithm based on Deep Deterministic Policy Gradient, referred to as E-DDPG. E-DDPG jointly optimizes sampling patterns, image reconstruction quality, and diagnostic accuracy. We introduces three key innovations: (1) a composite reward function that simultaneously encourages cross-entropy reduction, structural similarity improvement, and decrease in predictive entropy; (2) a percentile-based replay buffer that diversifies learning by equally sampling low- and high-value transitions; and (3) integration of the Straight-Through Gumbel-Softmax mechanism to preserve end-to-end differentiability while enabling discrete action selection. We evaluate E-DDPG against state-of-the-art RL-based methods and ablation variants on the fastMRI/fastMRI+ knee datasets at acceleration factors of 4X, 8X, and 10X, demonstrating its superior performance and validating the effectiveness of each proposed component.

Primary Area: reinforcement learning

Submission Number: 17325