Performance of deep learning-based segmentation of soft tissue sarcoma by MRI sequence, tumor type and location

Area of Science:

• Medical imaging
• Artificial intelligence in oncology
• Surgical planning

Background:

• Soft tissue sarcomas (STS) are rare and diverse tumors.
• Accurate segmentation of STS on preoperative MRI is crucial for surgical planning.
• Deep learning (DL) offers potential for automated segmentation.

Purpose of the Study:

• To evaluate the accuracy of a DL model for automated STS segmentation on preoperative MRI.
• To assess the impact of different MRI sequences, anatomical locations, and histological subtypes on model performance.

Main Methods:

• Retrospective analysis of 299 STS patient MRIs (2004-2022).
• Manual segmentation of tumors on fat-suppressed T1-weighted and T2-weighted sequences.
• Training and evaluation of separate 3D nnU-Net models for each sequence and combination using Dice, F2 score, ASSD, and HD95 metrics.

Main Results:

• Single-sequence T1 models outperformed multi-modal approaches for STS segmentation.
• The T1 axial model showed best volumetric accuracy (F2 0.91), while T1 sagittal excelled in boundary delineation (ASSD 2.1 mm, HD95 4.3 mm).
• Optimal sequence varied by anatomical site (T1 for extremities, T2 for abdomen/pelvis), and histology (highest for myxofibrosarcoma, lowest for leiomyosarcoma).

Conclusions:

• Deep learning models demonstrate high accuracy for soft tissue sarcoma segmentation.
• Model performance is significantly influenced by tumor location and histological subtype.
• T1-weighted MRI sequences are often superior for STS segmentation, and multi-modal fusion did not consistently improve results.