Learning From Partially Labeled Data for Multi-Organ and Tumor Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image segmentation is crucial for diagnosis and treatment planning.
• Partially labeled datasets pose challenges due to high costs of labor and expertise.
• Current methods often use single-task networks, limiting performance and increasing computational costs.

Purpose of the Study:

• To develop a versatile network capable of segmenting multiple organs and tumors from partially labeled medical image datasets.
• To address the limitations of single-task networks in terms of performance and computational efficiency.
• To introduce a novel Transformer-based dynamic on-demand network (TransDoDNet).

Main Methods:

• Proposed TransDoDNet, a hybrid network combining convolutional neural networks and Transformers.
• Implemented a dynamic head that adaptively generates kernels using Transformer's self-attention mechanism.
• Developed a large-scale partially labeled Multi-Organ and Tumor Segmentation (MOTS) benchmark.

Main Results:

• TransDoDNet demonstrated superior performance on seven organ and tumor segmentation tasks compared to competitors.
• The dynamic head effectively modeled long-range organ-wise dependencies.
• A pre-trained general 3D medical image segmentation model showed advanced performance over self-supervised methods.

Conclusions:

• TransDoDNet offers an efficient and flexible solution for multi-task medical image segmentation with partially labeled data.
• The proposed dynamic head and Transformer integration enable adaptive kernel generation and improved performance.
• The MOTS benchmark and pre-trained model provide valuable resources for advancing 3D medical image segmentation research.