Menu
The pipeline of the proposed multi-task evidential deep neural network for predicting the cell counts. The network consists of one encoder and three task-specific decoders. For the encoder, we utilize the pre-trained inceptionResNet v2 trained with ImageNet weights to capture imaging features, represented as latent features Z. The class-specific density map estimation decoder predicts K density maps for each cell type in the given image. The cell counting decoder predicts the cell counts, and the cell segmentation network segments the nuclei. We train the model end-to-end with a novel multi-task evidential learning loss function to predict and count various cell types. Cell type counting decoder is the main task, and class-aware density maps and nuclei segmentation are the
auxiliary tasks. The network output the predicted cell counts. For illustration purpose, we overlayed the predicted cell types on the original whole slide image. The color contours denote specific cell type. In this example network output, we presented the epithelial, connective and inflammatory cells in red, blue and green colors.
Predicting cell type counts in histology images can be a challenging task due to the complex and variable nature of tissue samples. In order to accurately predict cell counts, a machine learning model would need to be trained on a large and diverse dataset of histology images, and be able to accurately identify and classify different types of cells.
It is possible to develop a multi-task evidential deep neural network for predicting cell counts. This type of network would be trained to predict the counts of multiple types of cells simultaneously, using evidence from the input data to make more accurate predictions. The use of an evidential approach, which combines multiple predictions and their associated uncertainties, can improve the reliability of the cell count predictions. Deep neural networks have been shown to be effective at handling a wide range of prediction tasks, and have the potential to outperform traditional machine learning methods in certain cases. However, it is important to carefully evaluate the performance of the network and ensure that it is reliable and accurate before using it for decision making.
Our AI research team has developed a deep learning model using evidential multi-task learning to count cell types in histology images.
R. Gudhe, H. Behravan, M. Sudah, V.-M Kosmaa, and A. Mannermaa. “Predicting cell type counts in whole slide histology images using evidential multi-task learning”. SPIE Medical Imaging, 2023.
If you are interested in a project related to cell type count estimation from medical images, you can consider reaching out to us.
University of Eastern Finland, Institute of Clinical Medicine
Yliopistonranta 1 C, Canthia building, Kuopio, Finland
hamid.behravan[@]uef.fi
+358 50 5139836
Copyright 2023 – Hamid Behravan
