Deep Reinforcement Learning Driven Weakly Supervised Lesion Segmentation in Lung CT Sequence Images

Accurate and automated segmentation of interstitial lung disease (ILD) lesions from computed tomography (CT) sequence images is essential for quantitatively assessing disease progression and evaluating potential therapeutic strategies. The heterogeneity of ILD lesions and the complexity of manual annotations present significant challenges in achieving precise segmentation. Leveraging the inherent continuity of slice data in CT, where lesion shapes and locations exhibit systematic and progressive changes, we propose a deep reinforcement learning (DRL)-driven weakly supervised ILD lesion segmentation framework, called CT smoother agent (CTSA). Specifically, given an initial mask trained by an arbitrary weakly supervised network, CTSA uses the DRL algorithm to ensure the continuity of the CT sequence by changing the value of pixels in the optical flow (OF) change region, so as to obtain an optimized mask. Subsequently, CTSA uses a dual U-Net to train the constraint segmentation module based on the optimized mask and the initial mask. This process is cycled to achieve the purpose of segmenting ILD lesions. The policy network and segmentation network interact to optimize ILD lesion boundaries and segmentation results simultaneously. We evaluated our model on a dataset comprising 306 ILD patients, and the proposed CTSA improves dice similarity coefficient (DSC) by 9.3%, mean intersection over union (mIoU) by 21.7%, and Hausdorff distance (HD) by 1.522. The experimental results show that our proposed CTSA optimization model can achieve the state-of-the-art (SOTA) performance by learning the contextual information of CT sequences. Our code has been released at: https://anonymous.4open.science/r/CT-Smoother-Agent-1E8A