Interactive Lung Tumor Segmentation Using Denoising Diffusion Models

Denoising diffusion models (DDMs) have demonstrated strong generative capabilities for realistic image generation, and their ability to encode uncertainty has motivated applications for medical imaging such as tumor segmentation. Training and inference of denoising diffusion probabilistic models (DDPMs) on whole 3D medical scans is memory-intensive and computationally slow. Prior studies have addressed this challenge using accelerated sampling strategies such as denoising diffusion implicit models ( DDIM) and latent space compression approaches such as latent diffusion models, demonstrating improvements in computational efficiency. However, these methods still exhibit high inference times for full 3D volumes, posing a significant challenge for time-sensitive clinical workflows where rapid turnaround is essential. We have developed a novel human-assisted DDM approach (HA-DDM) to address this problem that integrates human guidance for interactive and fast medical image segmentation. Our method takes a user-defined point approximately at the center of the structure to be segmented as an input prompt and crops the image scan to a 3D bounding box region of interest (ROI) of a pre-determined size around the structure, centered at the point. This is fed to the DDPM and DDIM to obtain the output 3D segmentation, which is pre-trained using bounding box ROIs of similar size around the tumor extracted from the training data. We sample multiple segmentations using the DDPM and DDIM to generate spatial uncertainty maps. These segmentations are then ensembled to create the results which are resampled into the full 3D scan space. Our approach improves memory and computational efficiency of DDMs by processing smaller 3D bounding boxes instead of full 3D scans. In this work, we demonstrated our approach for lung tumor segmentation on 2D slices and 3D ROI bounding boxes. We validated our method on held-out test sets, achieving median Dice scores of 0.73 and 0.92 on the Medical Decathlon and institutional 2D datasets respectively using DDPM. For 3D volumes, we achieved Dice scores of 0.77 using DDIM on the institutional dataset with just 20 denoising steps instead of the standard 1000 in DDPM, significantly reducing computational time and resources without sacrificing segmentation quality.