After After
Before Before

10170 - IHC, Nuclei Detection, AI

DEVELOPED FOR AUTOMATIC SEGMENTATION OF NUCLEI IN IHC IMAGES

Identification and segmentation of individual nuclei in IHC images is of interest in many applications. Nuclei can be difficult to detect accurately and precisely using traditional image analysis with feature engineering. This APP utilizes artificial intelligence (AI) for automatic and robust nuclei segmentation in IHC stained tissue.

The nuclei detection performance has been assessed using precision and recall. Precision measures the ability of the APP to only detect viable nuclei and not mistake e.g. artefacts for nuclei. Recall (also known as sensitivity) measures the ability of the APP to detect all nuclei present. Precision and recall are combined to a F1-score which reflects both in a harmonic mean. Testing on 2,776 manually detected nuclei stained for ER, PR, or Ki-67, all three performance measures show excellent nuclei detection as seen from the table below.

The nuclei segmentation performance has been assessed using aggregated Jaccard index (AJI) [1] which captures the segmentation quality. Testing on 2,776 manually segmented nuclei, the performance measure shows excellent nuclei segmentation as seen from the table below.

 PrecisionRecall (Sensitivity)F1-score     AJI     
Mean 95.4 % 95.9 % 0.957 0.804


With an accurate nuclei segmentation, the APP can be configured to fulfill the needs of the user with Visiopharm’s Author™ module. The APP currently has two configuration options which can be switched on and off as needed:

  • Classification of DAB positive and negative nuclei to compute the positive percentage

  • Classification of DAB positivity (negative, 1+, 2+, 3+) to compute H-score

It is also possible for the user to customize the APP by adding post-processing steps and/or output variables.

KEYWORDS
Immunohistochemistry, IHC, Nuclei Segmentation, Nuclei Detection, Artificial Intelligence, AI, Digital Pathology

METHODS
The APP uses a deep learning network, which has been trained using 13,434 annotated nuclei from IHC stained breast cancer tissue (ER, PR, and Ki-67), to segment nuclei. The architectural structure of the network is a U-Net which is popular for medical image segmentation. The neural network uses a cascade of layers of nonlinear processing units for feature extraction and transformation, with each successive layer using the output from the previous layers as input. U-Net uses an encoder-decoder structure with a contracting path and an expansive path. For more information on the network architecture, see [2].

QUANTITATIVE OUTPUT VARIABLES
The APP can output the following variables:

  • Total Nuclei (#): The total number of nuclei

  • Negative Nuclei (#): The number of negative nuclei

  • Positive Nuclei (#): The number of positive nuclei

  • Positive Percentage (%): The percentage of positive cells among the total number of cells

  • H-Score [0-300]

  • Intensity Score [0-3]: A score between 0-3 depending on the mean DAB intensity of the positive nuclei

AUXILIARY APPs
There are no auxiliary APPs

WORKFLOW
Step 1: Load an IHC image
Step 2: Outline tumor either manually or automatically using one of Visiopharm’s solutions
Step 3: Load and run the APP “10170 – IHC, Nuclei Detection, AI”

STAINING PROTOCOL
There is no staining protocol available.

ADDITIONAL INFORMATION
To run the APP, a NVIDIA GPU with minimum 4 GB RAM is required. The APP utilizes the Visiopharm Engine™, Engine™ AI and Viewer software modules, where Engine™ and Engine™ AI offer an execution platform to expand processing capability and speed of image analysis. The Viewer allows a fast review together with several types of image adjustment properties e.g. outlining of regions, annotations and direct measures of distance, curve length, radius, etc. By adding the Author™ and/or Author™ AI modules the APP can be customized to fit other purposes. These modules offer a comprehensive and dedicated set of tools for creating new fit-for-purpose analysis APPs, and no programming experience is required.

REFERENCES

LITERATURE
1. Kumar, N., et al. A Dataset and a Technique for Generalized Nuclear Segmentation for Computational Pathology, IEEE Transactions on Medical Imaging, 2017, 36 (17), 1550-1560, DOI

2. Ronneberger, O., et al. U-Net: Convolutional Networks for Biomedical Image Segmentation, International Conference on Medical Image Computing and Computer-Assisted Intervention, 2015, 234-241, DOI

RUO
FIGURE 1
FIGURE 1
Breast cancer tissue stained with ER.
FIGURE 2
FIGURE 2
Outlining of tumor regions in breast cancer tissue stained with ER. Outlining can be done manually, or it can be done automatically using APP “10162 – IHC, Tumor Detection, AI”, as in this case, or using our virtual double staining technique.
FIGURE 3
FIGURE 3
Result of applying APP “10170 – IHC, Nuclei Detection, AI” to the outlined tumor regions. Automatic nuclei segmentation is achieved using AI and classification into negative/positive is based on the intensity of DAB staining in each nucleus.