Statistical considerations in spectral histopathology and performance evaluation
Shachi Mittal, Rohit Bhargava · 2021
Spectroscopic imaging offers a potential path to all-digital molecular pathology by relating the spectral data to histopathological details of tissue. In the different statistical approaches applied to spectral data, there are gaps in the appropriate sample size estimation for a significant statistical power and accurate model assessment especially for multiclass problems. Underestimation of the sample size can lead to statistically insignificant diagnostic tests while an overestimation can greatly increase experimental costs and time frames.Since the receiver operating characteristic (ROC) curve is designed primarily for a binary test, there are no straightforward approaches to use it for multiple classes in a typical pathology image. In this study, we have described sample size estimation (power analysis) and multiclass diagnostic ROC curve generation for hyperspectral datasets.