Deep Learning for Automated Detection of Cellular Profiles in Whole-slide Immunohistochemical Images


  • Islam Alzoubi
  • Rong Zhang
  • Guoqing Bao
  • Christina Loh
  • Svetlana Cherepanoff
  • Michael Buckland
  • Xiuying Wang
  • Manuel B. Graeber


Background: Routine examination of complete histological slides at cellular or subcellular resolution poses an insurmountable challenge for human observers. However, the availability of such high-resolution data on the distribution of proteins in tissues, e.g., obtained through immunohistochemical staining procedures, is highly desirable. Based on our previous work that established PathoFusion, we have explored different deep learning models, e.g., a bifocal convolutional neural network (BCNN), a residual network (ResNet), and ensemble methods for the automated detection of cells of interest that are immunoreactive for specific markers in brain tumour biopsies. 

Aims: The work aims to explore the capability of deep learning models to identify pathological changes at cellular resolution and to detect immunolabelled cells of interest.

Methods: A total of 50,741 patches of size 64×64 and 128 x 128 pixels were extracted from 36 whole-slide immunohistochemical images based on two types of expert-annotated coordinates indicating the presence of cells of interest and containing regions lacking such cells. We explored the effectiveness of the BCNN and its subnet structure [1] as well as state-of-the-art ResNet-50 [2] convolutional and ensemble methods, respectively. Using the PathoFusion framework, we converted the detection signals into corresponding heatmaps denoting the distribution of cells of interest in whole-slide images. 

Results: Our preliminary experimental results demonstrate that the built-in model (BCNN) of the PathoFusion framework achieved consistent results compared with ResNet-50 in terms of accuracy (~97%) and F1-score (0.97), while the subnet structure achieved slightly lower performance. When ResNet-50 was ensembled with the subnet through model averaging, the performance was improved, i.e., some detection noise was removed. Quantitative comparisons and whole-slide cross-modality analyses are still being studied.  




Our preliminary experiments indicate that PathoFusion is suitable for the autonomous detection of cellular and perhaps subcellular structures. Recent state-of-the-art deep learning strategies will be investigated to improve PathoFusion further in order to allow working in different microscopy and other image analysis scenarios.