Tissue-Background Segmentation


Logo for Tissue-Background Segmentation

About

Image Version:
54699642-ffcd-4fb1-9a30-6fdb938c3478
Last updated:
May 21, 2019, 6:23 p.m.

Interfaces

This algorithm implements all of the following input-output combinations:

Inputs Outputs
1
  • Generic Medical Image (Image)
  • Generic Overlay (Heat Map)
  • Results JSON File (Anything)
  • Model Facts

    Summary

    Tissue-background segmentation of histopathological whole-slide images. In this project we collected 100 + 8 whole-slide images (WSIs) for development ant testing of a fully-convolutional neural network (FCNN) that can distinguish between tissue and backgound areas on the WSIs.

    Mechanism

    This model is a fully-convolutional neural network consisting of 7 convolutional layers with ReLU (Maas et al., 2013) activation function in the first 6 convolutional layers and softmax in the last one. Max pooling was inserted after each of the first 3 convolutional layers to reduce the memory requirements of the network.

    The model was trained with WSI in tiff format of breast, lymph node, kidney, tongue, rectum, and lung tissue. The slides were stained with 6 different stains: hematoxylin and eosin (H&E), Sirius Red, Periodic Acid-Schiff (PAS), cytokeratin AE1/AE3 (AE1AE3), Ki-67, and a cocktail of cytokeratin 8 and cytokeratin 18 (CK8-18).

    Validation and Performance

    The model was validated with a dataset containing lung, cornea, aorta, brain, skin, uterus, and kidney tissue samples. The tissues were stained with Grocott, Alcian Blue, Von Kossa, Perls, and Chromotrope Aniline Blue (CAB) stains. Only lung and kidney tissues were present in both of the datasets, while the stains are non-overlapping.

    Uses and Directions

    This algorithm was developed for research purposes only.

    Warnings

    Common Error Messages

    Information on this algorithm has been provided by the Algorithm Editors, following the Model Facts labels guidelines from Sendak, M.P., Gao, M., Brajer, N. et al. Presenting machine learning model information to clinical end users with model facts labels. npj Digit. Med. 3, 41 (2020). 10.1038/s41746-020-0253-3