TCNN: A Transformer Convolutional Neural Network for artifact classification in whole slide images

Abstract

The production of pathological slides is a complex task requiring several physical and chemical procedures that are often done manually. Occasionally, such procedures may end up in patterns (called artifacts) appearing in the images that were not present in the biological tissue but are produced during the slide processing. This paper presents a new method for the automatic detection of such artifacts in pathological images. This detection is seen as a binary classification task that is solved by means of a new Transformer Convolutional Neural Network (TCNN). The proposed method can assist laboratory technicians in avoiding the time-consuming manual labeling of images, preventing the risk of sending artifact patches to pathologists and physicians. The artifact patches include misleading data and should not be considered. In addition, artifact patches jeopardize the accuracy of Computer-Aided Diagnosis (CAD) systems. The proposed method works on the Hue-Saturation-Value colormap layer, which is related to color representation, to improve model robustness against external lighting changes. Spatial Transformer layers are used to make the model spatially invariant on input data, and a set of convolutional layers of EfficientNet Convolutional Neural Network improve pattern recognition. Images used in this research are more than 13,000 and represent 22 types of tissue affected by 7 diverse types of artifacts. Through performance measurement, we evaluate the model using several metrics: Probability Calibration graph, Mean Prediction Probability, Brier Loss, ROC-AUC curve, Confusion Matrix, Accuracy, Sensitivity, Specificity, F1-Score, and Confidence Interval.