A neural pathomics framework for classifying colorectal cancer histopathology images based on wavelet multi-scale texture analysis

Abstract

Colorectal cancer (CRC) constitutes the third most commonly diagnosed cancer in males and the second in females. Precise histopathological classification of CRC tissue pathology is the cornerstone not only for diagnosis but also for patients' management decision making. An automated system able to accurately classify different CRC tissue regions may increase diagnostic precision and alleviate clinical workload. However, tissue classification is a challenging task due to the variability in morphological and textural characteristics present in histopathology images. In this study, an artificial neural network was trained to classify between eight classes of CRC tissue image patches derived from a public dataset with 5000 CRC histopathology image tiles. A total of 532 multi-level pathomics features examined at different scales were extracted by visual descriptors such as local binary patterns, wavelet transforms and Gabor filters. An exhaustive evaluation involving a variety of wavelet families and parameters was performed in order to shed light on the impact of scale on pathomics based CRC tissue differentiation. Our model achieved a performance accuracy of 95.3% with tenfold cross validation demonstrating superior performance compared to 87.4% reported in recent studies. Furthermore, we experimentally showed that the first and the second levels of the wavelet approximations can be used without compromising classification performance.

Figure 1

Performance analysis of wavelet families: Each line shows the percentage of each mother wavelet with respect to the total number of models achieving different levels of accuracy. db daubechies, sym symlets, coif coiflets, bior biorthogonal, rbio reverse biorthogonal. (Figure created in Excel 2013, https://www.microsoft.com/microsoft-365/excel).

Figure 2

Performance analysis among different levels of wavelet packet transform: the first and second levels exhibit high accuracies while in higher levels of decomposition the texture analysis error is significantly affected. (Figure created in Excel 2013, https://www.microsoft.com/microsoft-365/excel).

Figure 3

Samples from the examined dataset and the proposed multi-level textural analysis pipeline featuring wavelet packet transform, local binary patterns, Gabor, grey level co-occurrence matrices, first order statistics concatenated in a pathomics signature for support vector machines and artificial neural networks tissue classification. (Figure created in PowerPoint 2013, https://www.microsoft.com/microsoft-365/powerpoint).

Figure 4

The proposed multi-level feature extraction strategy including several combinations of WPT, Gabor filters, LBP, GLCM, first and high order statistics. WPT wavelet packet transform, LBP local binary patterns, GLCM grey level co-occurrence matrices, FOS first order statistics, HOS higher order statistics. (Figure created in PowerPoint 2013, https://www.microsoft.com/microsoft-365/powerpoint).

Figure 5

Differences among mother wavelets at the second level of decomposition for part (red bounding box) of the original image. db daubechies, sym symlets, coif coiflets, bior biorthogonal, rbio reverse biorthogonal. (Figure created in PowerPoint 2013, https://www.microsoft.com/microsoft-365/powerpoint).