2017/05/02 by Abhijit Guha Roy, Sailesh Conjeti, Roy, Abhijit Guha +9 · 2 citations
Computer Science · Medicine · #Advanced Neural Network Applications #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Medical Image Segmentation Techniques #Radiomics and Machine Learning in Medical Imaging
paper · pdf · doi:10.48550/arxiv.1705.00938
openalex publication_date 2017/05/02 · openalex created_date 2022/10/04 · openalex updated_date 2026/07/28
Training deep fully convolutional neural networks (F-CNNs) for semantic image\nsegmentation requires access to abundant labeled data. While large datasets of\nunlabeled image data are available in medical applications, access to manually\nlabeled data is very limited. We propose to automatically create auxiliary\nlabels on initially unlabeled data with existing tools and to use them for\npre-training. For the subsequent fine-tuning of the network with manually\nlabeled data, we introduce error corrective boosting (ECB), which emphasizes\nparameter updates on classes with lower accuracy. Furthermore, we introduce\nSkipDeconv-Net (SD-Net), a new F-CNN architecture for brain segmentation that\ncombines skip connections with the unpooling strategy for upsampling. The\nSD-Net addresses challenges of severe class imbalance and errors along\nboundaries. With application to whole-brain MRI T1 scan segmentation, we\ngenerate auxiliary labels on a large dataset with FreeSurfer and fine-tune on\ntwo datasets with manual annotations. Our results show that the inclusion of\nauxiliary labels and ECB yields significant improvements. SD-Net segments a 3D\nscan in 7 secs in comparison to 30 hours for the closest multi-atlas\nsegmentation method, while reaching similar performance. It also outperforms\nthe latest state-of-the-art F-CNN models.\n