Duodenal segmentation model for celiac disease pathology
Hirvonen, Vilho (2024)
2024
Bioteknologian ja biolääketieteen tekniikan maisteriohjelma - Master's Programme in Biotechnology and Biomedical Engineering
Lääketieteen ja terveysteknologian tiedekunta - Faculty of Medicine and Health Technology
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2024-12-16
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2024121010936
https://urn.fi/URN:NBN:fi:tuni-2024121010936
Tiivistelmä
Background and aims:
Celiac disease is an autoimmune disorder caused by ingestion of gluten. In genetically pre-disposed individuals the immune system recognizes complexes formed by partially degraded gluten peptides and specific HLA haplotypes. The duodenal mucosa is damaged by this process, and this can be observed histologically as villus blunting and crypt elongation. These features form the basis of histological evaluation metrics such as the Marsh-Oberhuber grading.
The only currently approved treatment for celiac disease is strict adherence to the gluten-free diet. This diet is not only difficult to adhere to, but it is also ineffective in individuals with non-responsive celiac disease. There are many medical trials underway which aim to improve the quality of life of patients by reducing the dependence of the gluten free diet. Histological assessment is a necessary validation metric in these trials, as it best describes long-term patient outcomes. However, these assessment methods suffer from poor inter-observer agreement.
Supervised deep learning methods such as convolutional neural networks have been used to assist pathologists in various applications. Networks with semantic segmentation outputs give pixel-accurate information about scanned whole slide images. To date, there have been only a few celiac disease focused studies which utilize segmentation models, which is partly due to limited data availability. Supervised deep learning methods require large training datasets which can be produced manually, but this annotation process is slow and subject to bias. However, it is possible to automatically generate histological datasets by using immunohistochemical staining techniques and image engineering methods.
Methods:
Histological dataset for identifying duodenal villus and crypt epithelium was generated by re-staining hematoxylin-eosin-stained slides with immunohistochemistry targeting villus and crypt epithelium. U-Net convolutional network was optimized with this data to output villus and crypt epithelial regions from hematoxylin-eosin-stained regions of interest. The generated model was validated with images from a separate test-set. The ratio of epithelial areas was used as a validation metric for the semantic segmentation model.
Results and conclusions:
The U-Net model’s crypt to villus epithelial ratio correlated with pathologist defined villus height to crypt depth ratio in the test set. The epithelial ratio was also converted to a categorical scale and compared with visually evaluated Marsh-Oberhuber scores. Comparisons with both categorical and continuous scales had high correlations. This model improved pathologist agreement and is now in routine use at Jilab, the site of work.
Celiac disease is an autoimmune disorder caused by ingestion of gluten. In genetically pre-disposed individuals the immune system recognizes complexes formed by partially degraded gluten peptides and specific HLA haplotypes. The duodenal mucosa is damaged by this process, and this can be observed histologically as villus blunting and crypt elongation. These features form the basis of histological evaluation metrics such as the Marsh-Oberhuber grading.
The only currently approved treatment for celiac disease is strict adherence to the gluten-free diet. This diet is not only difficult to adhere to, but it is also ineffective in individuals with non-responsive celiac disease. There are many medical trials underway which aim to improve the quality of life of patients by reducing the dependence of the gluten free diet. Histological assessment is a necessary validation metric in these trials, as it best describes long-term patient outcomes. However, these assessment methods suffer from poor inter-observer agreement.
Supervised deep learning methods such as convolutional neural networks have been used to assist pathologists in various applications. Networks with semantic segmentation outputs give pixel-accurate information about scanned whole slide images. To date, there have been only a few celiac disease focused studies which utilize segmentation models, which is partly due to limited data availability. Supervised deep learning methods require large training datasets which can be produced manually, but this annotation process is slow and subject to bias. However, it is possible to automatically generate histological datasets by using immunohistochemical staining techniques and image engineering methods.
Methods:
Histological dataset for identifying duodenal villus and crypt epithelium was generated by re-staining hematoxylin-eosin-stained slides with immunohistochemistry targeting villus and crypt epithelium. U-Net convolutional network was optimized with this data to output villus and crypt epithelial regions from hematoxylin-eosin-stained regions of interest. The generated model was validated with images from a separate test-set. The ratio of epithelial areas was used as a validation metric for the semantic segmentation model.
Results and conclusions:
The U-Net model’s crypt to villus epithelial ratio correlated with pathologist defined villus height to crypt depth ratio in the test set. The epithelial ratio was also converted to a categorical scale and compared with visually evaluated Marsh-Oberhuber scores. Comparisons with both categorical and continuous scales had high correlations. This model improved pathologist agreement and is now in routine use at Jilab, the site of work.