ImageXD presentation 30 March 2017. Developing software for biological image analysis using classic compute vision techniques and looking forward to deep learning segmentation and classification.
5. • Reproducible results
• Minimal computer vision knowledge
• Biology-focused documentation
• Modular and extensible
• Run locally or distributed
Good stuff
6. • Difficult to process bright-field data
• Support for proprietary microscope formats
• Pipeline configuration takes time
• 3D algorithms assume isotropic data
• Computationally expensive
• Results are “good enough”
Not-so-good stuff
10. • Address usability issues
• Transition from methods that generalize across
images to methods that generalize across
experiments, or even assays and microscopes.
What’s next?
11. • Address usability issues
• Transition from methods that generalize across
images to methods that generalize across
experiments, or even assays and microscopes.
• Transition from image segmentation (pixel labeling) to
semantic segmentation (scene labeling)
What’s next?