Spatial assessments in texture analysis: what the radiologist needs to know.

Area of Science:

• Radiomics and Medical Imaging
• Machine Learning in Oncology

Background:

• Radiomics studies often involve extensive feature extraction, particularly texture analysis for intratumoral heterogeneity.
• Texture analysis methods include spatial filters and neighborhood-based approaches to capture image variations.

Purpose of the Study:

• To highlight the importance of appropriate methodologies in radiomics-based predictive modeling.
• To address challenges related to high dimensionality and performance evaluation in radiomics.

Main Methods:

• Investigating spatial texture assessments for intratumoral heterogeneity.
• Discussing data dimensionality reduction techniques and their correct application to training data.
• Examining performance metrics, including receiver operating characteristic curves and confusion matrices.

Main Results:

• Spatial texture features are effective for assessing intratumoral heterogeneity in oncologic imaging.
• Dimensionality reduction must be applied cautiously to training data to prevent overfitting.
• Confusion matrices offer clinically significant insights, especially for unbalanced datasets.

Conclusions:

• Optimizing radiomics models requires careful feature selection and robust validation strategies.
• Accurate assessment of intratumoral heterogeneity through radiomics can correlate with tumor biology.
• Appropriate performance evaluation is key for translating radiomics findings into clinical practice.