Pizza and Lecture: U-Net Deep Learning Strategies for Slice-wise Segmentation of Tomographic 3D Datasets

#southern #Minnesota #deep #learning #u-net #models #tomographic #3D #data

An automated and generally applicable method for segmentation is still a focus of medical image processing research. Artificial inteligence methods have shown promising results in recent years, especially with widely available scalable Deep Learning libraries. In this work, a five layer hybrid U-net is developed for slice-by-slice segmentation of liver data sets. Training data is taken from the Medical Segmentation Decathlon database, providing 131 fully segmented volumes. A slice-oriented segmentation model is implemented utilizing deep learning algorithms with adaptions for variable parenchyma shape along the stacking direction and similarities between adjacent slices. Both are transformed for coronal and sagittal views. The implementation is on a GPU rack with TensorFlow and Keras. For a quantitative measure of segmentation accuracy, standardized volume and surface metrics are used. Results DSC=97.59, JI=95.29 and NSD=99.37 show proper segmentation comparable to 3D U-Nets and other state of the art. The development of a 2D-slice oriented segmentation is justified by short training time and less complexity and therefore massively reduced memory consumption.

We further evaluated how the U-net cascade can be applied to different medical domains with less available reference data for training, more precisely MRI data of the liver from VIBE acquisition protocol. With Graph cuts as a post-processing tool, expert knowledge can overcome the black-box nature of Deep Learning models in a semi-automated way. Weights for combining original image edges and edges from the DL segmentations result from Evolution strategy optimization, closing the gap in accuracy of Deep Learning and Graph cut in general but still conserving the potential for expert-driven corrections. Although the U-net cascade lacks a sufficient amount of MRI training data leading to mediocre accuracy, Graph cuts is a powerful tool for post-processing and iterative enrichment of the training basement. Another strategy for enriching small reference datasets are generative adversial networks (GANs). It is shown that datasets enriched by synthesized slices improve the training process of Deep Learning models.

• Date: 20 Jan 2020
• Time: 06:30 PM to 08:00 PM
• All times are (UTC-06:00) Central Time (US & Canada)
• Add Event to Calendar
  iCal
  Google Calendar

• 321 3rd Ave SW
• Rochester, Minnesota
• United States 55902
• Building: Medical Sciences Bldg (east side, north door)
• Room Number: Mann Hall

---

• Contact Event Host
• Co-sponsored by Victoria Marks

• Starts 09 January 2020 11:03 AM
• Ends 19 January 2020 05:00 PM
• All times are (UTC-06:00) Central Time (US & Canada)
• No Admission Charge

• 

A pizza social will begin at 6:30 PM, followed by the start of the talk at 7:00 PM.