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

SIMPLE SUMMARY: One of the most important and hardest tasks during cancer diagnostics and operations is to differentiate between cancerous and non-malignant tissues. It is important to detect and remove all cancerous cells, but at the same time cut as little as possible non-malignant tissue. That is why there is now a high demand for methods that will help to recognize margins of cancer. One non-invasive, new and promising method is hyperspectral imaging (HSI) with the use of machine learning (ML). However, the output of the ML models is usually fuzzy, especially when pixel-wise classification methods are used. This research shows that, (1) the impact of particular pre-processing techniques on the performance of tissue recognition depends on the ML model used and its architecture and could be explained by the wavelengths emphasized by the models, and (2) that the application of post-processing can strongly improve performance (both sensitivity and specificity) and the ability to easily differentiate between tissue types. ABSTRACT: Background: Recent studies have shown that hyperspectral imaging (HSI) combined with neural networks can detect colorectal cancer. Usually, different pre-processing techniques (e.g., wavelength selection and scaling, smoothing, denoising) are analyzed in detail to achieve a well-trained network. The impact of post-processing was studied less. Methods: We tested the following methods: (1) Two pre-processing techniques (Standardization and Normalization), with (2) Two 3D-CNN models: Inception-based and RemoteSensing (RS)-based, with (3) Two post-processing algorithms based on median filter: one applies a median filter to a raw predictions map, the other applies the filter to the predictions map after adopting a discrimination threshold. These approaches were evaluated on a dataset that contains ex vivo hyperspectral (HS) colorectal cancer records of 56 patients. Results: (1) Inception-based models perform better than RS-based, with the best results being 92% sensitivity and 94% specificity; (2) Inception-based models perform better with Normalization, RS-based with Standardization; (3) Our outcomes show that the post-processing step improves sensitivity and specificity by 6.6% in total. It was also found that both post-processing algorithms have the same effect, and this behavior was explained. Conclusion: HSI combined with tissue classification algorithms is a promising diagnostic approach whose performance can be additionally improved by the application of the right combination of pre- and post-processing.