Assessment of discriminant models in infrared imaging using constrained repeated random sampling - Cross validation

Area of Science:

• Biomedical imaging
• Spectroscopy
• Data analysis

Background:

• Infrared (IR) imaging is a powerful, label-free technique for analyzing molecular composition in cells and tissues.
• Current methods using multivariate analysis (e.g., Partial Least Squares-Discriminant Analysis) on IR images can lose spatial information.
• Internal validation methods like k-fold cross-validation may yield overly optimistic results when pixel size is below spatial resolution.

Purpose of the Study:

• To address the overestimation of model performance in IR image analysis.
• To introduce a novel, unbiased internal validation method for IR microscopy data.
• To improve the accuracy and robustness of discriminant models in IR-based tissue and cell analysis.

Main Methods:

• Evidence of overly optimistic internal validation (repeated k-fold cross-validation) for IR images with small pixel sizes.
• Proposal and description of COnstrained Repeated Random Subsampling-Cross Validation (CORRS-CV) for unbiased model evaluation.
• CORRS-CV utilizes constrained random sampling of training pixels without replacement to generate training and test subsets.

Main Results:

• Demonstrated that standard cross-validation can be unreliable for IR images with sub-resolution pixels.
• CORRS-CV effectively circumvents oversampling issues, leading to more accurate model performance estimation.
• The proposed method ensures more robust and reliable discriminant models for IR microscopy.

Conclusions:

• Standard internal validation methods can be misleading in IR imaging analysis.
• CORRS-CV offers a more accurate and robust approach for evaluating discriminant models in IR microscopy.
• This method enhances the analysis of biochemical differences in cells and tissues using IR imaging.