‘MEDICAL IMAGENET’

Petteri Teikari, PhDversion Sun 5 February 2017

Creating database of volumetric biomedical images facilitatingdeep learning image analysis for niche fields with no ‘big data’as images needed are reduced

IMAGENET

Petteri Teikari, PhDversion Sun 5 February 2017

Benchmark database with the dogs and all used for the developmentof better deep learning architectures

Medical ImageNetTransfer learning for biomedical image analysis

Andrej Karpathy: What I learned from competing against a ConvNet on ImageNet Sep 2, 2014

http://karpathy.github.io/2014/09/02/what-i-learned-from-competing-against-a-convnet-on-imagenet/

Imagenet: A large-scale hierarchical image database [PDF] stanford.edu

J Deng, W Dong, R Socher, LJ Li, K Li… - Computer Vision and …, 2009 – ieeexplore.ieee.orgAbstract: The explosion of image data on the Internet has the potential to foster more sophisticated and robust models and algorithms to index, retrieve, organize and interact with images and multimedia data. But exactly how such data can be harnessed and organized

Cited by 3488 Related articles All 30 versions

Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

K He, X Zhang, S Ren, J Sun - Proceedings of the IEEE …, 2015 – cv-foundation.org Cited by 769

ResNet from Microsoft then beat Andrej Karpathy:

IMAGENET for ‘Transfer Learning’

Petteri Teikari, PhDversion Sun 5 February 2017

In practice, leverage big general database on specific image analysistask such as medical analysis

Medical ImageNetTransfer learning for biomedical image analysis

Yosinski et al. (2014)( Cited by 413):

“Many deep neural networks trained on natural images exhibit a curious phenomenon in common: on the first layer they learn features similar to Gabor filters and color blobs. Such first-layer features appear not to be specific to a particular dataset or task, but general in that they are applicable to many datasets and tasks.”

http://dx.doi.org/10.1109/TMI.2016.2535302

https://arxiv.org/abs/1608.08614

Given the same number of training classes, is it better to have coarse classes or fine-grained classes? Which is better: more classes or more examples per class?

https://arxiv.org/abs/1701.06599“Effect of reduction in the training data on the performance of CNNs trained from scratch and deeply fine-tuned CNNs.”

‘MEDICAL IMAGENET’ for Transfer Learning

Petteri Teikari, PhDversion Sun 5 February 2017

Very little volumetric data available in the end compared to ImageNet’snumber for natural RGB images.

Medical ImageNetTransfer learning for biomedical image analysis

Medical Image Analysis Volume 36, February 2017, Pages 61–78http://dx.doi.org/10.1016/j.media.2016.10.004

Clinical practice and research, MRI, CT, PET, OCT, ultrasound, etc.

Fundamental research, light and electron microscopy, MRI, OCT, ultrasound, etc.

Example deep learning volumetric project with code sharing still too uncommon in neuroscience / clinical image analysis.

Open source tools for large-scale neuroscienceJ Freeman - Current opinion in neurobiology, 2015

‘MEDICAL IMAGENET’ for Transfer Learning

Petteri Teikari, PhDversion Sun 5 February 2017

In practice: Start compiling a database with various modalities for example as classes, and start maybe ‘fine-graining’ later with different pathologies per modality

Medical ImageNetTransfer learning for biomedical image analysis

Three-dimensional (3D) Retinal OCT

Three-photon microscopymouse brain vasculature

Quantifying focused ultrasound brain stimulation with MRI (and EEG)

A 3D ultrasound imaging transducer generates a volumetric image of a fetus

Holographic optical-phase microscope with AI reconstruction

A colorized electron microscope image of the Ebola virus 

2-photon imaging of hippocampal neurons

Volume rendering of the abdominal cavity from a CT angiography examination, revealing a large abdominal aortic aneurysm. 

Spine MRI for back pain investigation

‘MEDICAL IMAGENET’ Data Sharing in Practice

Petteri Teikari, PhDversion Sun 5 February 2017

For the image database, even some stacks unused for final publications are useful in regard to image analysis uses, and probably even easier for researchers to shareas that would not give their results to the hands of their competitors

Medical ImageNetTransfer learning for biomedical image analysis

DeepMedichttps://biomedia.doc.ic.ac.uk/software/deepmedic/

1st Layer 2nd Layer 3rd Layer

For example, the 1st layer weights could be ‘universal’ [see. e.g. Yosinski et al. (2014)] for various fields from MRI to microscopy, and only the higher layers would be fine-tuned to be domain-specific (see e.g.

TensorFlow tutorial)

‘MEDICAL IMAGENET’ Summary

Petteri Teikari, PhDversion Sun 5 February 2017

VISION: Democratize deep learning for various clinical and fundamental research fields with limited available datasets, both for high-tech and low-tech imaging techniques.

Medical ImageNetTransfer learning for biomedical image analysis

http://dx.doi.org/10.1007/978-3-319-46976-8_20

Fei-Fei Li https://www.ted.com/talks/fei_fei_li_how_we_re_teaching_computers_to_understand_pictures

http://www.forbes.com/sites/quora/2016/12/28/why-is-it-important-to-democratize-machine-learning/#67b9c6394908Answer by François Chollet, Deep learning researcher at Google, author of Keras

https://blog.keras.io/building-powerful-image-classification-models-using-very-little-data.html