Deep Learning tissue type from examples

Deep learning using Convolutional Neural Networks is the state of the art technique for recognising tissue type based on visual appearance. Deep Learning harnesses the massive power of NVIDIA GPUs to build models that replicate the way the human brain recognises objects in a hierarchical way, building up application dependent features from the bottom up. Such methods have been shown to outperform traditional tissue recognition based on manually crafted features (texture, cell detections, etc.) by a considerable margin. The HG-MIM Deep Learning add-on allows you to harness the power of deep learning within the user friendly web-based environment of HG-MIM. Features include:

• Bespoke HeteroGenius convolutional networks specifically designed to process large colour Digital Pathology slides accurately and efficiently.
• Learn models based on relatively few annotated example areas that need not be precisely drawn. No need to laboriously annotate whole images to build models.
• Automatic annotation of tissue areas - Create editable annotations.
• Automatic Quantification - Generate high resolution overlay images, area and perimeter based statistics for whole slides and multiple annotated sub-regions. Fully integrated with the HG-MIM Pro reports generation system.
• Automatic Stereology - Integration with HG-MIM Pro stereology add-on allowing points or regions to be automatically assigned a tissue type (tag), with easy comparison tohuman markup.
• Colour normalisation and data transforms make learned models more transferrable to unseen data with stain varying from that seen in the training examples.
• Easy install, with no GPU configuration required. Automatically detects your Graphics Card and uses it. Falls back to CPU if no GPU present (not recommended).
• Validate the accuracy (and kappa, precision, recall etc.) of learned models based on further annotations.
• Integration with HG-MIM 3D pathology add-on allowing classified volumetric data sets and surfaces to be generated (see below).