Classification of colorectal tissue histopathological images using amalgamated hand-crafted features

Histopathological inspection is a frequently utilized method to detect and diagnose colorectal cancer (CC). The existing methods for colorectal tissue classification using histopathological images (HIs) utilize spatial information. Spectral analysis offers valuable insights into the frequency components and noise characteristics of images, thereby facilitating more effective processing and detection techniques. However, the absence of spectral information in existing methods leads to moderate performance in classifying colorectal tissue HIs. To improve the performance of colorectal tissue classification using HIs, this work proposes a set of amalgamated hand-crafted features formulated from the same image by combining histogram of oriented gradients (HOG), scale-invariant feature transform (SIFT), and cepstrum domain features. The developed amalgamated hand-crafted features incorporate both spatial and spectral information, providing a more comprehensive representation for the accurate classification of colorectal tissue HIs. Further, the developed features are given as input to a proposed deep neural network (DNN). In addition, a modified softmax function is proposed to attain better classification performance. The softmax function is updated during the training process. The effectiveness of the proposed network is compared with the other existing current approaches. From the experimental analysis, the proposed method attained an accuracy of 94.24%, sensitivity of 94.12%, precision of 94.32%, specificity of 99.20%, and F1 score of 0.9422 to classify colorectal tissue. The proposed method has the potential to enhance the accuracy and reliability of clinical diagnosis assessments. • A non-local mean (NLM) filter is utilized to denoise histopathological images (HIs). • NLM filter has efficiently reduced noise while retaining the edges of HIs. • Amalgamated hand-crafted features are developed using SIFT, HOG, and cepstrum domain features. • The developed hand-crafted feature uses spatial as well as spectral information. • A modified softmax function is proposed to attain better classification performance.