Impact of Pre- and Post-Processing Steps for Supervised Classification of Colorectal Cancer in Hyperspectral Images

Area of Science:

• Medical imaging
• Computational pathology
• Cancer diagnostics

Background:

• Hyperspectral imaging (HSI) and neural networks are emerging tools for colorectal cancer detection.
• Pre-processing techniques are crucial for optimizing neural network performance in HSI analysis.
• The impact of post-processing on HSI-based cancer detection models remains less explored.

Purpose of the Study:

• To evaluate the impact of pre-processing and post-processing techniques on hyperspectral imaging-based colorectal cancer detection.
• To compare the performance of Inception-based and RemoteSensing (RS)-based 3D-CNN models.
• To assess the effectiveness of median filter-based post-processing algorithms.

Main Methods:

• Two pre-processing techniques (Standardization, Normalization) were applied.
• Two 3D-CNN models (Inception-based, RS-based) were utilized.
• Two median filter-based post-processing algorithms were tested on ex vivo hyperspectral colorectal cancer data from 56 patients.

Main Results:

• Inception-based models outperformed RS-based models, achieving 92% sensitivity and 94% specificity.
• Model performance varied with pre-processing: Inception-based models favored Normalization, while RS-based models favored Standardization.
• Post-processing improved overall sensitivity and specificity by 6.6%, with both algorithms yielding similar effects.

Conclusions:

• Hyperspectral imaging with tissue classification algorithms offers a promising diagnostic approach for colorectal cancer.
• Optimizing the combination of pre- and post-processing techniques can further enhance diagnostic performance.
• The study highlights the significant role of post-processing in improving HSI-based cancer detection accuracy.