<![CDATA[Histopathological examination of Haematoxylin and Eosin (H&E) stained tissues is the foundation of diagnosing colorectal cancer (CRC). The increasing digitization of pathology has resulted in a surfeit of images. Currently, computational methods to automatically classify histopathological tissues are in their infancy. In this paper, nine-class colorectal tissue classification was evaluated using EfficientNet-B0, ResNet50, and DenseNet121. The classification problem was standardized with respect to data pre-processing, data augmentation, model optimization and evaluation. The classification models were developed using the NCT-CRC-HE-100K dataset (100,000 histopathology image patches) and validated on the independent CRC-VAL-HE-7K dataset (7,180 image patches). The validation was performed using accuracy, macro-average precision, macro-average recall, macro-F1 score, weighted-F1, and class-wise average precision, average recall, average F1 score and average specificity. Class-level analysis revealed that MUS and STR were challenging classes for the studied architectures with DEB resulting in confusion. A qualitative Grad-CAM analysis was performed on representative images to understand the reason for the predicted class. The study revealed that the three CNN architectures were able to generalize well for colorectal histopathology classification. Controlled external evaluation of deep learning models is essential to understand the class-specific reasons for the errors.]]>
