BA/MA: Feature Visualization for Skin Lesions
Advisor: Federico Tombari
Supervision by:Fernando Navarro
Deep neural networks have proven to give an outstanding performance in the classification task. However, understanding the learned features and the learning process is still a challenging task. In this project, we aim to investigate and visualize the relevant learned features that networks learn in the context of skin lesion classification. The project consists of developing software tools to assist dermatologist in the diagnosis of a skin lesion by providing both the classification result and a visualization of relevant feature for the diagnosis.
Figure 1: Project Description. Bring together feature visualization techniques in computer vision and skin lesion classification aiming for better diagnosis in the field of dermoscopy
Background and Motivation
Skin cancer is the most common type of cancer. Therefore, there is a need for computer-aided systems to classify and detect cancerous skin lesion. Deep neural networks have proven successful results when a large annotated dataset is available. Nevertheless, for skin lesion classification, the amount of data is scale-prohibitive. To overcome this problem, webly supervised learning has previously been investigated giving a notable improvement compared to training with a limited annotated dataset . On the other hand, understanding how and what the networks have learned is still a challenging task. Many works on this direction have been proposed. In , they proposed a taxonomy for feature visualization, which classifies methods into three categories: input modification methods, deconvolutional methods, and Input reconstruction methods. In medical applications, there are few works in this direction. Example of feature visualization are femur fractures  and breast cancer . However, the visualization of learned features in skin lesion classification is still a novel research topic.
The aim of this project is to investigate which features are relevant for skin lesion classification in the ten class categorization. The student would compare the learned features from several neural networks trained on 1) a small annotated dataset and 2) a large web-crawled dataset in a webly supervised strategy.
- Implementation of a feature visualization for skin lesions in the same fashion as .
- Visualize the learned features for a small but well-annotated dataset of skin lesions in different state-of-the-art networks.
- Visualize the learned features for a large web-crawled dataset of skin lesions in different state-of-the-art networks
- Experimental evaluation from the different developed systems.
Depending on the desired thesis type and student's preferences, the tasks will be adapted to suit the scope of the thesis.
- Knowledge of image processing
- Experience with deep learning
- Good python programming
- Experience with deep learning frameworks like TensorFlow?
 Chen, X. and Gupta, A.: Webly supervised learning of convolutional networks. In ICCV, 2015
 Grün, F., Rupprecht, C., Navab, N. and Tombari, F., 2016. A taxonomy and library for visualizing learned features in convolutional neural networks. arXiv preprint arXiv:1606.07757.
 Kirchhoff, S., Biberthaler, P., Navab, N. and Mateus, D., 2017, September. Automatic Classification of Proximal Femur Fractures Based on Attention Models. In Machine Learning in Medical Imaging: 8th International Workshop, MLMI 2017, Held in Conjunction with MICCAI 2017, Quebec City, QC, Canada, September 10, 2017, Proceedings (Vol. 10541, p. 70). Springer.
 Geçer, B., 2016. Detection and classification of breast cancer in whole slide histopathology images using deep convolutional