Performance of Machine Learning Methods Based on Multi-Sequence Textural Parameters Using Magnetic Resonance Imaging...

Area of Science:

• Radiology
• Machine Learning
• Oncology

Background:

• Soft tissue tumors require accurate differentiation between malignant and benign types for appropriate patient management.
• Multiparametric magnetic resonance imaging (mpMRI) offers rich data for tumor characterization.

Purpose of the Study:

• To evaluate a machine learning (ML) method for differentiating malignant from benign soft tissue tumors.
• To assess the diagnostic performance of ML models using mpMRI textural features and clinical information.

Main Methods:

• 163 patients with pathologically confirmed soft tissue tumors underwent mpMRI.
• Twelve histographic and textural parameters were extracted from various MRI sequences.
• Machine learning models (support vector machine) were developed using textural features, clinical data, or a combination.
• Model performance was evaluated using area under the receiver operating characteristic curves (AUC) via fivefold cross-validation.

Main Results:

• The combined ML model incorporating textural features and clinical information achieved the highest diagnostic ability (AUC 0.89).
• The clinical information model (AUC 0.85) showed non-inferior performance compared to models using only textural features (AUC 0.79-0.84).
• The combined model's performance was comparable to that of two experienced radiologists (AUCs 0.89 and 0.87).

Conclusions:

• Machine learning methods utilizing mpMRI textural features and clinical data provide robust diagnostic performance.
• This approach aids in differentiating malignant from benign soft tissue tumors effectively.
• The combined ML model demonstrates potential as a valuable tool in soft tissue tumor diagnosis.