Evaluation of Preprocessing Methods on Independent Medical Hyperspectral Databases to Improve Analysis

Area of Science:

• Medical imaging
• Computational pathology
• Cancer diagnostics

Background:

• Cancer remains a leading global cause of mortality, necessitating advancements in early and accurate detection.
• Hyperspectral Imaging (HSI) combined with artificial intelligence (AI) shows promise for cancer detection.
• Standardized preprocessing methods for medical HSI data are lacking, hindering algorithm performance.

Purpose of the Study:

• To evaluate and compare various preprocessing techniques for medical HSI data.
• To determine the optimal preprocessing strategy for enhancing tumor detection across different cancer types.
• To assess the impact of preprocessing on AI-based pixel-wise classification of HSI data.

Main Methods:

• Evaluated combinations of spatial/spectral smoothing, Min-Max scaling, Standard Normal Variate (SNV) normalization, and median spatial smoothing.
• Utilized two machine learning and deep learning models for pixel-wise classification.
• Tested preprocessing methods on HSI datasets for colorectal, esophagogastric, and brain cancers.

Main Results:

• Preprocessing significantly impacts tumor identification performance.
• Median Filter preprocessing yielded slightly better results for colorectal tumors (AUC 0.94).
• Min-Max scaling preprocessing achieved higher accuracy for esophagogastric (AUC 0.93) and brain tumors (AUC 0.92).

Conclusions:

• Min-Max scaling is identified as the most relevant and robust preprocessing technique for HSI-based cancer detection across diverse datasets and instrumentation.
• Median Filter can smooth critical spectral features, leading to performance variability.
• Standardized preprocessing is essential for reliable AI-driven HSI cancer diagnostics.