Abstract
Medical image segmentation is challenging due to the need for expert annotations and the variability of these manually created labels. Previous methods tackling label variability focus on 2D segmentation and single modalities, but reliable 3D multi-modal approaches are necessary for clinical applications such as in oncology. In this paper, we propose a framework for generating reliable and unbiased labels with minimal radiologist input for supervised 3D segmentation, reducing radiologists' efforts and variability in manual labeling. Our framework generates AI-assisted labels through a two-step process involving 3D multi-modal unsupervised segmentation based on feature clustering and semi-supervised refinement. These labels are then compared against traditional expert-generated labels in a downstream task consisting of 3D multi-modal bone tumor segmentation. Two 3D-Unet models are trained, one with manually created expert labels and the other with AI-assisted labels. Following this, a blind evaluation is performed on the segmentations of these two models to assess the reliability of training labels. The framework effectively generated accurate segmentation labels with minimal expert input, achieving state-of-the-art performance. The model trained with AI-assisted labels outperformed the baseline model in 61.67% of blind evaluations, indicating the enhancement of segmentation quality and demonstrating the potential of AI-assisted labeling to reduce radiologists' workload and improve label reliability for 3D multi-modal bone tumor segmentation. The code is available at https://github.com/acurtovilalta/3D_LabelGeneration .