Weakly-Supervised Anomaly Detection assisted by Attention Models

Thesis by: Amelia Jimenez Sanchez
Advisor: Prof. Nassir Navab
Supervision by: Shadi Albarqouni, Diana Mateus

Abstract

Localization of anatomical regions of interest (ROIs) is a natural pre-processing step in many medical image analysis tasks, for example in diagnosis. While it is sometimes trivial for physicians, it turns to be tedious and very time consuming for them. Convolutional Neural Networks (CNNs) have proven to be very successful in computer vision tasks as object detection and image classification, due to the ability of extracting rich and hierarchical features that are useful for both: localization and classification.

In this thesis, we investigate the concepts of Self-Transfer Learning (STL) and Spatial Transformer Network (STN). We want to exploit STL which jointly optimizes classification and localization only with weak labels, no localization information is provided, and STN that allows us to find a canonical representation by means of learning invariance to scale, rotation, translation and more generic warping. We want to explore whether the use of STN in combination with STL will improve the classification performance and if STL could assist in the anomaly detection task.

We evaluate our model using three medical image datasets, chest X-rays, femur X-rays and mammograms, and compare them to previous weakly supervised approaches.

Literature

Resultant Paper

2017
	A. Kazi, S. Albarqouni, A. Sanchez, C. Kirchhoff, P. Biberthaler, N. Navab, D. Mateus Automatic Classification of Proximal Femur Fractures based on Attention Models Proceedings of MICCAI Workshop on Machine Learning in Medical Imaging (MLMI), Quebec, Canada, September 2017 (bib)