Histopathological cancer images classification with Deng entropy

Histopathological imaging is of paramount importance for the initial detection, diagnosis, and classification of tumors. Recurrent neural networks and convolutional neural networks have led to substantial advancements in digital pathology, thereby enhancing classification accuracy. Tsallis and Shannon entropies were employed to optimize cancer classification. However, certain constraints remain to be addressed, including the need to mitigate noise and uncertainty. This study aimed to classify histopathological images of cancer using Deng entropy and long short-term memory (LSTM) as a novel approach that provides accurate assessments for pathologists to differentiate between normal and abnormal tissues. The computed Deng entropy, based on the box covering method, is used as a vector to feed bidirectional LSTM (bLSTM) networks and obtain Deng's information dimensions. This innovative approach obtains Deng entropy from different scales (box sizes), yielding a measure that captures differences in complexity. Three histopathological datasets (BreakHis, Lung-colon, and PANDA) were analyzed. Statistical tests were performed on each dataset to determine the most effective discrimination of histopathological images between the information and Deng information dimensions. The binary breast classification exhibited a higher performance accuracy rate of 0.98. For multiclass analysis, the accuracy was 0.99. The lung image model exceeded a classification accuracy of 0.98, and for the colon, it was 0.99. The prostate image accuracy was 0.924. Deng entropy provides a precise classification system for histopathological images of breast, colon, and lung cancers. Our results demonstrated that the proposed methodology can achieve satisfactory cancer classification.