Development and Validation of Multi-Omics Thymoma Risk Classification Model Based on Transfer Learning

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Thymoma risk stratification is crucial for determining appropriate treatment strategies.
• Accurate noninvasive methods for predicting thymoma risk are needed to optimize patient management.
• Current methods may not fully leverage the potential of integrated clinical and imaging data.

Purpose of the Study:

• To develop and evaluate a prediction model integrating clinical, radiomics, and deep features for stratifying thymoma risk.
• To assess the performance of a fusion model compared to individual models (clinical, radiomics, deep features) using transfer learning.
• To determine the efficacy of the developed model in noninvasively distinguishing between high and low-risk thymoma.

Main Methods:

• A cohort of 150 thymoma patients was retrospectively analyzed, with data split into training (80%) and testing (20%) sets.
• Clinical, radiomics (2590 features), and deep features (192 features) were extracted from CT images (non-enhanced, arterial, venous phases).
• Feature selection was performed using ANOVA, Pearson correlation, PCA, and LASSO, followed by SVM classification and transfer learning for model development.

Main Results:

• The fusion model achieved high performance, with AUCs of 0.99 (training) and 0.95 (test), and accuracies of 0.93 (training) and 0.83 (test).
• Compared to individual models, the fusion model demonstrated superior performance in risk stratification.
• Clinical, radiomics, and deep models showed varying degrees of success, but were outperformed by the integrated approach.

Conclusions:

• The fusion model integrating clinical, radiomics, and deep features is effective for noninvasively stratifying thymoma risk.
• This approach shows significant potential to aid in determining optimal surgical strategies for thymoma patients.
• Transfer learning enhances the model's ability to accurately predict thymoma risk, offering a valuable tool in clinical oncology.