This repository is for our work submitted to MICCAI24, titled "Comprehensive Generative Replay for Task-Incremental Segmentation with Concurrent Appearance and Semantic Forgetting".
In this work, we synthesize paired images and segmentation masks to simulate past task data, by employing a Bayesian Joint Diffusion (BJD) model to preserve image-mask correspondence, and equipping a Task-Oriented Adapter (TOA) on the prompt embedding to modulate the diffusion model for task-scalable data synthesis. When encountering new tasks, we leverage these replayed past task data to evoke faded memory, and update it to incorporate new task knowledge for future replays.
Prepare three segmentation tasks, such as Cardiac, Fundus, and Prostate.
First, run train_bjdwithtoa.sh to learn three tasks sequentially to simulate image-mask pair of each task. The training hyper-parameters can be set in the code. Notably, the pre-trained weights of CLIP text encoder can be downloaded from CLIP text encoder. Then, run train_segnetwork.py to update the segmentation network sequentially. Similarly, the training hyper-parameters can be set in the options\base_options.py.