Computer Vision News - November 2021

52 AI Research Paper SOTA baselines were used for the evaluation in the problem statement of novelty detection and the training data were taken as “anomaly free” to examine two challenging medical problems with different image characteristics and appearance of abnormalities: metastases detection in digital pathology and chest X-Rays. The task is to detect metastases in H&E stained images of lymph nodes in the Camelyon16 challenge by training anomaly detection models only on healthy tissue aiming to identify tissue with metastases. Tissues exhibiting metastasis may differ from healthy types only by texture, spatial structure, or distribution of nuclei. Chest X-ray is one of the most common examinations for diagnosing various lung diseases and the task was to recognize 14 findings, such as Atelectasis or Cardiomegaly, on the chest X-rays in the NIH dataset (ChestX-ray14 dataset). To create the training dataset, the existing one was split into two parts with only posteroanterior (PA) or anteroposterior (AP) projections. This is an example from the Camelyon 16 challenge (top) and the NIH dataset (bottom). Next to the image the anomaly prediction score by the proposed method. The higher score signifies an anomaly. It’s interesting to note that borderline cases are correctly evaluated by the model. The autoencoder architecture for image anomaly detection, where g is the network, f the feature extractor, x^ the reconstructed image while the reconstruction L calculates the difference between and