Development and External Validation of Deep-Learning-Based Tumor Grading Models in Soft-Tissue Sarcoma Patients Using...

Area of Science:

• Radiology
• Oncology
• Artificial Intelligence

Background:

• Soft-tissue sarcoma (STS) tumor grading is crucial for treatment planning.
• Current grading relies on invasive pathological analysis of biopsies.
• Deep learning (DL) offers a potential non-invasive imaging analysis method for STS characterization.

Purpose of the Study:

• To develop and evaluate DL models for non-invasive differentiation of low-grade (G1) and high-grade (G2/G3) STS.
• To assess the performance of DL models based on MRI sequences for tumor grading.

Main Methods:

• Retrospective analysis of contrast-enhanced T1-weighted fat-saturated (T1FSGd) and fat-saturated T2-weighted (T2FS) MRI sequences from two cohorts (training: 148 patients, testing: 158 patients).
• Development of DL models using transfer learning with the DenseNet 161 architecture.
• Tumor grading determined by French Federation of Cancer Centers Sarcoma Group criteria on pre-therapeutic biopsies.

Main Results:

• T1FSGd and T2FS DL models achieved AUCs of 0.75 and 0.76, respectively, on the test cohort.
• The T1FSGd model yielded the highest F1-score (0.90).
• The T2FS model demonstrated significant risk stratification for overall survival; attention maps highlighted key tumor features.

Conclusions:

• MRI-based DL models can predict STS tumor grading with good reproducibility in external validation.
• These models offer a promising non-invasive tool for STS grading and risk stratification.