Colorectal Tissue Image Classification Across Datasets

This work studies the performance of the pretrained models to classify colorectal tissues and the effectiveness of tissue image classification models when it is applied across different datasets that may exhibit different color characteristics. In particular, the study is conducted on three pre-trained CNN models, ResNet50, VGG19, and EfficientNetB3, where they are used as to extract a feature vector of the image patch. Then the multi-layer perceptron-based neural network is used to perform the task of multiclass classification. The result shows that the best accuracy to classify the colorectal histopathological images was achieved by the ResNet model with 93.87% test accuracy on the Kather-texture2016 image dataset. And for the accuracy of the classification model when it is trained in the public dataset and applied to the local datasets, the normalized color image test dataset showed the best accuracy rate at 80.69% when compared to raw images and grayscale images among the local datasets. This result suggests that the pre-trained classification models could be useful for tissue image classification tasks across various laboratory sources and image color adjustment could help enhance the recognition performance even further of the image classification model. Therefore, consideration for image color correction would be needed to preserve the recognition performance of the model if the pre-trained tissue image classification model is to be used off-the-shelf across laboratories.