Ensemble Building of State-of-the-art Models for Skin Lesion Boundary Segmentation
Ensemble of FCN, SegNet, and DeepLabv3+ models for automated skin lesion boundary segmentation in dermoscopic images.
Abstract
This manuscript summarizes our method and validation re- sults for the ISIC Challenge 2018 Skin Lesion Analysis Towards Melanoma Detection Task 1: Lesion Boundary Segmentation. Aim of this task is to develop an approach that automatically segments skin lesions within dermoscopic images. Our convolutional neural network based approaches utilize pre-trained weights for the FCN, SegNet, and DeepLabv3+ (i.e. with the Xception network backbone) as initialization. After training our networks on 90% of all skin lesion images and building an ensemble out of four best performing models, our approach yields a mean thresholded jaccard-index of 76.0% for the ISIC task 1 validation dataset. The single best performing model achieved 80.2%.
Ivo Matteo Baltruschat
Co-Founder / Research Scientist
Ivo Matteo Baltruschat studied Information and Electrical Engineering at the University of Applied Sciences, Hamburg between 2010 and 2014. In 2016, he finished his Master of Science in medical engineering science at the Universität zu Lübeck with his thesis on Deep Learning for Advanced Medical Applications. During DAISYlabs, he was a PhD student in the group of Tobias Knopp at the Institute for Biomedical Imaging, Hamburg University of Technology. His research covered the automatic analysis of medical x-ray images using machine learning methods, with a focus on applying deep learning to high-resolution medical x-ray images for computer-aided diagnosis.
Thilo Sentker
Research Scientist
Frederic Madesta
PhD Student
Rüdiger Schmitz
PhD Student
Helge Kniep
PhD Student
Nils Gessert
Research Scientist
René Werner
Co-Founder / Head of scientific working group
René Werner co-founded DAISYlabs and led the scientific working group at the University Medical Center Hamburg-Eppendorf, coordinating deep learning research for biomedical image processing across UKE and TUHH.
