Uncertainty estimation in the classification of histopathological images with HER2 overexpression using Monte Carlo Dropout

Abstract

Classification of breast cancer tissues with HER2 overexpression into 0, 1+, 2+, or 3+ categories is a crucial clinical task for determining HER2 positivity, and hence, prescribing the adjuvant HER2-targeted therapy. Deep learning-based methods have been fundamental for creating support tools that improve decision-making, and they have provided outstanding performance in the classification of histopathological images. However, the quantification of the uncertainty has been traditionally neglected when deep learning-based methods are used, despite its critical importance in healthcare applications. In this work we propose a new method, using deep learning and the Monte Carlo Dropout, to measure the uncertainty when classifying breast cancer tissue images with HER2 overexpression.The proposed method has four main stages including (1) WSI pre-processing, (2) patch dataset generation, (3) Bayesian deep learning-based classification, (4) and prediction and uncertainty estimation. For efficient computation of the predictive distribution, we propose a 2-step predictive distribution method that decreases significantly the execution time. Furthermore, the method achieves tissue-level classification by training the classifier on a patch-level and using aggregation techniques. The proposed method achieved on average 0.89 accuracy, 0.81 precision, and 0.74 recall for classification on a whole slide tissue-level. The method is also capable of characterizing the higher and lower uncertainty regions in the whole slide tissue images by estimating the predictive entropy and mutual information. This additional piece of information can be very important for practitioners to take decisions in the healthcare domain.