Minimally Interactive Segmentation of Soft-Tissue Tumors on CT and MRI using Deep Learning

Douwe J. Spaanderman,Martijn P. A. Starmans,Gonnie C. M. van Erp,David F. Hanff,Judith H. Sluijter,Anne-Rose W. Schut,Geert J. L. H. van Leenders,Cornelis Verhoef,Dirk J. Grunhagen,Wiro J. Niessen,Jacob J. Visser,Stefan Klein

2024-02-13

Abstract:Segmentations are crucial in medical imaging to obtain morphological, volumetric, and radiomics biomarkers. Manual segmentation is accurate but not feasible in the radiologist's clinical workflow, while automatic segmentation generally obtains sub-par performance. We therefore developed a minimally interactive deep learning-based segmentation method for soft-tissue tumors (STTs) on CT and MRI. The method requires the user to click six points near the tumor's extreme boundaries. These six points are transformed into a distance map and serve, with the image, as input for a Convolutional Neural Network. For training and validation, a multicenter dataset containing 514 patients and nine STT types in seven anatomical locations was used, resulting in a Dice Similarity Coefficient (DSC) of 0.85$\pm$0.11 (mean $\pm$ standard deviation (SD)) for CT and 0.84$\pm$0.12 for T1-weighted MRI, when compared to manual segmentations made by expert radiologists. Next, the method was externally validated on a dataset including five unseen STT phenotypes in extremities, achieving 0.81$\pm$0.08 for CT, 0.84$\pm$0.09 for T1-weighted MRI, and 0.88\pm0.08 for previously unseen T2-weighted fat-saturated (FS) MRI. In conclusion, our minimally interactive segmentation method effectively segments different types of STTs on CT and MRI, with robust generalization to previously unseen phenotypes and imaging modalities.

Image and Video Processing,Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The paper aims to address the segmentation problem of soft tissue tumors (STT) in CT and MRI images. Specifically, the paper develops a deep learning-based semi-interactive segmentation method to improve segmentation accuracy and efficiency.

### Main Objectives:
1. **Improve Segmentation Accuracy**: By having the user click six points near the tumor boundary, the deep learning model is assisted in segmentation, thereby improving accuracy.
2. **Enhance Segmentation Efficiency**: Compared to fully manual segmentation, this method significantly reduces the workload of doctors and offers higher accuracy and robustness than fully automatic segmentation methods.
3. **Enhance Generalization Ability**: This method can achieve good generalization performance on different types of STT and different imaging modalities, including unseen tumor types and imaging sequences.

### Problems Addressed:
- **Manual Segmentation is Time-Consuming and Observer-Dependent**: Current manual segmentation methods, while accurate, are time-consuming and highly dependent on the observer.
- **Poor Performance of Fully Automatic Segmentation**: Fully automatic segmentation methods usually perform poorly in soft tissue tumor segmentation, especially when dealing with diverse tumor types and imaging modalities.

### Experimental Results:
- On the WORC test dataset, the semi-interactive segmentation method achieved an average DSC of 0.85±0.11 on CT and 0.84±0.12 on T1-weighted MRI, both higher than fully automatic segmentation methods.
- On the external validation dataset TCIA, the method also demonstrated good generalization performance on unseen tumor types and imaging sequences (such as T2-weighted fat-suppressed MRI).

In summary, the method proposed in this paper accurately segmented different types of soft tissue tumors across various imaging modalities and demonstrated good generalization ability.