Clinical Scorecard: Automated Segmentation of Soft-Tissue Tumors in CT and MRI Using Deep Learning Techniques

At a Glance

Key Highlights

• Manual segmentation of STTs is time-consuming, costly, and observer-dependent, necessitating automated methods.
• InteractiveNet uses six user-defined interior margin points to guide segmentation, improving generalizability and efficiency.
• The method integrates best practices from the nnU-Net framework and was validated on heterogeneous public datasets including WORC and TCIA.

Guideline-Based Recommendations

Diagnosis

• Use CT or T1-weighted and T2-weighted MRI for imaging STTs.
• Employ manual reference segmentations by experienced musculoskeletal radiologists for ground truth.

Management

• Apply minimally interactive deep learning segmentation methods to assist tumor delineation.
• Use six interior margin points to define volume of interest and generate exponentialized geodesic distance maps to guide segmentation.
• Incorporate state-of-the-art CNN architectures such as nnU-Net for improved accuracy.

Monitoring & Follow-up

• Evaluate segmentation accuracy against manual reference segmentations on independent datasets.
• Use volume and diameter measurements derived from segmentations for clinical assessment.

Risks

• Fully automatic segmentation methods may perform sub-optimally due to STT heterogeneity.
• Insufficient user interaction points may lead to incomplete tumor encapsulation and inaccurate segmentation.

Patient & Prescribing Data

Patients with soft-tissue tumors undergoing imaging for diagnosis or treatment planning

Minimally interactive segmentation reduces physician workload and improves segmentation accuracy compared to fully automatic methods, facilitating clinical decision-making.

Clinical Best Practices

• Obtain six interior margin points in three orthogonal planes to define tumor boundaries accurately.
• Use exponentialized geodesic distance maps to enhance tumor voxel identification based on intensity and spatial differences.
• Integrate preprocessing, network architecture optimization, training, post-processing, and ensembling from nnU-Net framework for robust segmentation.
• Validate segmentation models on heterogeneous, multi-center datasets to ensure generalizability.

References

• WORC Database for STT Imaging
• The Cancer Imaging Archive (TCIA) STT Dataset
• Luo et al. Interactive Segmentation Framework
• nnU-Net Framework for Medical Image Segmentation