Efficient Human-in-the-Loop Pancreatic Tumor Annotation via Large-Scale Pre-Trained Model with Adaptive Post-Processing

Deep learning has significantly impacted fields like medical image analysis. However, the effectiveness of supervised learning models depends on large volumes of annotated data. Annotating medical images, especially for tumor regions, is costly and time-intensive. To overcome this limitation, this paper propose a novel framework to accelerate tumor annotation, producing a unique dataset with pancreatic lesions labeled in 300 CT scans from multiple centers. Traditional annotation methods would require an experienced radiologist around 216 hours for this task, whereas our framework completed it in 57 hours, achieving similar or even improved annotation quality. This success stems from two core elements: (1) pre-training the model on a large-scale multi-organ dataset, then fine-tuning it on the target domain before using the fine-tuned model to generate preliminary labels, and (2) adaptive post-processing based on statistical insights from an extensive in-house JHH dataset and clinical evidence. More importantly, our framework could help both the AI model and annotation processes, substantially reducing the annotation costs required to produce large-scale datasets for tumor detection and segmentation tasks for other organs, such as liver and kidney. With our framework, we introduce AbdomenAtlas2.0-Mini, a dataset of 300 CT scans with high-quality voxel-level pancreatic lesions annotations.