Automated Deep Learning Segmentation of Soft-Tissue Tumors in CT and MRI

Overview

This study developed a minimally interactive deep learning method, InteractiveNet, for segmenting soft-tissue tumors (STTs) on CT and MRI. InteractiveNet demonstrated improved generalizability and accuracy compared to fully automatic and previous interactive methods, validated on heterogeneous public datasets.

Background

Soft-tissue tumors are rare and heterogeneous, occurring in various body locations with variable progression. Accurate 3D tumor segmentation is critical for radiotherapy planning, radiomics, and treatment response evaluation but is currently manual, time-consuming, and observer-dependent. Deep learning-based automatic segmentation has shown promise but struggles with STT heterogeneity and imaging modality variability. Incorporating minimal user interaction may enhance segmentation accuracy and clinical applicability.

Data Highlights

Key Findings

• InteractiveNet requires six user clicks near tumor boundaries to define a volume of interest and generate an exponentialized geodesic distance map to guide segmentation.
• Incorporation of nnU-Net best practices improved the original interactive segmentation framework's performance.
• InteractiveNet outperformed fully automatic nnU-Net and the original MIDeepSeg method in accuracy and generalizability on internal and external datasets.
• The method successfully segmented STTs across multiple imaging modalities (CT, T1-weighted MRI, T2-weighted FS MRI) and diverse tumor phenotypes.
• Volume and diameter measurements derived from InteractiveNet segmentations showed potential for clinical use in tumor assessment.

Clinical Implications

InteractiveNet offers a practical, time-efficient approach to STT segmentation by combining minimal user input with deep learning, reducing manual workload and observer variability. Its ability to generalize across tumor types and imaging modalities supports integration into clinical workflows for radiotherapy planning and quantitative imaging biomarker extraction. This method may enhance precision in tumor measurement and treatment response evaluation.

Conclusion

The minimally interactive deep learning segmentation method InteractiveNet provides accurate, generalizable STT delineations on CT and MRI, addressing limitations of fully automatic approaches. Its clinical adoption could streamline tumor segmentation and improve patient management.

References

1. Luo et al. 2021 -- Interactive Segmentation Framework for Medical Imaging
2. Isensee et al. 2021 -- nnU-Net: Self-configuring Method for Medical Image Segmentation
3. WORC Database -- Public Dataset for Soft-Tissue Tumors
4. TCIA -- The Cancer Imaging Archive Soft-Tissue Tumor Dataset