convolutional neural networks

Set of KNIME workflows for the training of a deep learning model for image-classification with custom images and classes.

The workflows take ground-truth category annotations as a table generated by the qualitative annotations plugins in Fiji.

Workflows for the training of a model AND for the prediction of image-category for new images are provided.

There are different workflows if you do:

- binary image-classification (images get classified in 1 category out of 2 possible categories) 

- classification from possibly more than 2 categories (images are classified in 1 category out of N possible categories).

The training workflows take care of image pre-processing and allows the visualization of the training and validation losses in real time along the training.  

For the training, transfer learning from a pre-trained VGG16 base is performed, with freshly initialized fully connected layers.

Only the fully connected layers are trained, the VGG16 base is frozen is this workflow, but once the fully connected layers trained the base could also be finetuned. In practice, it often works well with the frozen base.

Chainer is a Python-based deep learning framework aiming at flexibility. It provides automatic differentiation APIs based on the define-by-run approach (a.k.a. dynamic computational graphs) as well as object-oriented high-level APIs to build and train neural networks. It also supports CUDA/cuDNN using CuPy for high performance training and inference. For more details of Chainer, see the documents and resources listed above and join the community in Forum, Slack, and Twitter.

An ImageJ plugin for DEFCoN, the fluorescence spot counter based on fully convolutional neural networks

NiftyNet is a TensorFlow-based open-source convolutional neural networks (CNNs) platform for research in medical image analysis and image-guided therapy. NiftyNet’s modular structure is designed for sharing networks and pre-trained models. Using this modular structure you can:

• Get started with established pre-trained networks using built-in tools;
• Adapt existing networks to your imaging data;
• Quickly build new solutions to your own image analysis problems.