Machine Learning Model Integrating Computed Tomography Image-Derived Radiomics and Circulating miRNAs to Predict...

Area of Science:

• Oncology
• Radiology
• Molecular Diagnostics

Background:

• Metastatic nonseminomatous germ cell tumors (mNSGCTs) are primarily treated with chemotherapy, often leaving residual disease.
• Accurate noninvasive methods are crucial for determining the histology of post-chemotherapy residual masses.
• Distinguishing teratoma from non-teratoma is vital for guiding treatment and surgical planning.

Purpose of the Study:

• To enhance the accuracy of predicting teratoma histology in post-chemotherapy residual masses.
• To integrate computed tomography (CT) radiomics features with circulating microRNAs (miR371-375).
• To differentiate between teratoma and non-teratoma histology noninvasively.

Main Methods:

• Retrospective analysis of 111 lesions from post-chemo CT scans, segmented using 3D Slicer.
• Extraction of radiomics features and measurement of plasma miR371-375 levels via real-time PCR.
• Evaluation of machine learning models combining radiomics and miR371-375 for predictive accuracy.

Main Results:

• The CatBoost model integrating radiomics and miR371-375 (R + 371 + 375) demonstrated superior predictive accuracy (AUC 0.96 training, 0.83 testing).
• Radiomics signature was the strongest independent predictor of teratoma histology (P < .0001) in multivariate analysis.
• Clinical factors like AFP, hCG, initial pathology, and lymph node status were also significant predictors.

Conclusions:

• Combining miR371-375 levels with CT radiomics features significantly improves teratoma prediction accuracy.
• This integrated approach offers a promising noninvasive tool for guiding treatment decisions in mNSGCTs.
• Potential to optimize patient management by reserving surgery for teratoma-containing residual masses.