Research on imbalance machine learning methods for MRT 1WI soft tissue sarcoma data.

Area of Science:

• Oncology
• Medical Imaging
• Machine Learning

Background:

• Soft tissue sarcoma is a rare, heterogeneous tumor where pathological grading is crucial for prognosis and treatment.
• Clinical data for soft tissue sarcoma is often imbalanced, posing challenges for accurate classification.
• Accurate, non-invasive preoperative prediction of tumor grade is essential for patient management.

Purpose of the Study:

• To propose an effective machine learning solution for classifying imbalanced soft tissue sarcoma data.
• To identify the optimal imbalance machine learning model for predicting soft tissue sarcoma classification.
• To support the development of personalized treatment plans for soft tissue sarcoma patients.

Main Methods:

• Radiomics methods were used to extract features from T1-weighted (T1-WI) images.
• Exploration of feature selection, sampling, and classification techniques led to 17 imbalance machine learning models.
• The Recursive Feature Elimination (RFE) technique, SMOTETomek (STT) sampling, and Extremely Randomized Trees (ERT) classification were investigated.

Main Results:

• The combination of RFE, STT, and ERT (RFE+STT+ERT) demonstrated the best performance among the tested models.
• The RFE+STT+ERT model achieved an accuracy of 81.57%, comparable to biopsy.
• Using an alternative dataset splitting method, the RFE+STT+ERT model reached an accuracy of 95.69%.

Conclusions:

• The developed machine learning method (RFE+STT+ERT) effectively addresses the challenge of imbalanced data classification in soft tissue sarcoma.
• This approach offers a promising non-invasive tool for preoperative pathological grade prediction.
• The findings can significantly contribute to personalized treatment strategies for soft tissue sarcoma.