Ahmed Serag - Unsupervised Anomaly Detection: Application to Colorectal Liver Metastasis

Background & objectives

The building of deep learning models needs large amounts of data with annotated examples. Obtaining expert labels for such models is difficult since detailed annotation is timeconsuming. Here, we present an unsupervised learning approach capable of identifying anomalous patterns that can serve as imaging biomarker candidates. We show the application of the proposed approach for detecting colorectal metastases in liver biopsies.


We propose to use unsupervised learning to create a rich generative model of normal (i.e. non-tumor) local anatomical appearance. We use generative adversarial networks to solve the problem of creating an adequately representative model of appearance, while at the same time learning a generative and discriminative component. We use information from both image space and latent space to differentiate between observations that conform to the training data and such data that does not fit.


The trained model is able to generate images that are visually similar to the non-tumor images. In the case of anomalous images (contain metastases), the pairs of input images and generated images show significant differences. The distributions of the anomaly score over non-tumor images from the training set and test set or over images extracted from metastatic cases show that the anomaly score is suitable for the classification of normal and anomalous samples (P<0.001).


We presented an unsupervised approach for colorectal metastasis detection in liver biopsies. Training patches were extracted from non-tumor images; avoiding the necessity of having detailed pathologist annotation. The approach may be applied across the whole range of computational pathology problems.