Image-based Classification of Tumor Type and Growth Rate using Machine Learning: a preclinical study

Area of Science:

• Medical imaging analysis
• Machine learning in oncology
• Preclinical cancer research

Background:

• Magnetic resonance (MR) imaging is crucial for cancer detection and diagnosis.
• Machine learning can extract texture features from MR images for personalized treatment.
• Texture features from T1-weighted post-contrast scans offer potential for tumor classification and growth prediction.

Purpose of the Study:

• To evaluate texture features from T1-weighted post-contrast MR scans for classifying brain tumors.
• To predict tumor growth rate in a preclinical mouse model using texture features.
• To optimize machine learning models by varying gray-level co-occurrence matrix (GLCM) sizes, tumor region selection, and model types.

Main Methods:

• Utilized texture features derived from T1-weighted post-contrast MR images.
• Employed varying gray-level co-occurrence matrix (GLCM) sizes and tumor region selections.
• Tested different machine learning models, including random forest classification and a two-layer feedforward neural network.

Main Results:

• A random forest model with a GLCM size of 512 achieved high specificity (91-92%) and sensitivity (73-89%) for classifying GL261, U87, and Daoy tumors.
• Tenfold cross-validation showed 84% accuracy using entire tumor volume and 74% accuracy using central tumor regions for classification.
• A neural network predicted tumor growth with a 16% mean squared error.

Conclusions:

• Texture features from standard MR images can effectively classify brain tumor types.
• Predictive models utilizing these features can forecast tumor growth rates.
• Model performance is influenced by GLCM size, tumor region selection, and tumor type, highlighting the need for tailored approaches.