Improving the Accuracy of Intrahepatic Cholangiocarcinoma Subtype Classification by Hidden Class Detection via Label Smoothing

Obtaining ground-truth labels for supervised training is a labor-intensive and time-consuming task. Owning to their large size, only slide-level labels or a handful of coarse annotations are usually provided for pathology images, which makes the training of the classifier challenging. In this study, we propose a conceptually simple, two-stage approach to classify small and large duct types in intrahepatic chalongiocarcinoma using only slide-level labels. Unlike conventional pathology image analysis methods employ multiple instance learning (MTh) applied to overcome the problem of the slide-level label, we introduce a novel label smoothing method to progressively refine the training labels to improve the classification accuracy. The main idea is that we introduce the hidden class, which is assumed to be mutually inclusive of all ground-truth classes and less confident for classification. By iteratively refining (i.e., smoothing) per-patch labels, we can extract and discard the hidden class from the training data. We demonstrate that the proposed label filtering scheme improves the classification accuracy by up to 30% compared to the baseline MTh method and 10% compared to the state-of-the-art noisy label cleaning method. In addition, we demonstrate the effectiveness of gene mutation prior information in the classification of two different duct types. The experimental results suggest that the proposed method may provide pathologists insight into the study of correlations between genetic and histologic subtypes.