Artificial Intelligence in Oncology - Supporting scientific research
Radboudumc
In short, Chella van der Posts' research entails the following:
Within this project they will develop AI methods to refine DGC diagnostics. Since DGC may be easily missed or hard to find on biopsies and prophylactic hereditary gastrectomy specimens, AI will aid the pathologist in the diagnostic work-up, improving the detection of relevant cell types among a very large set of slides, with high potential to improve cancer diagnostics. Furthermore, the automation in cell detection provided by AI algorithms will allow to quantitatively and objectively assess DGC patterns in large series of slides, potentially giving new insights in specific morphological features of DGC, such as patterns of spatial cell distributions.
With this project, researchers aim at using AI to better identify and classify future patients with (H)DGC, to increase detection of individual patients and families, that might eventually result in better patient stratification for therapeutic options and clinical decisions. This will give more insight into specific features of CDH1 mutated DGC, both in a hereditary as well as sporadic setting. In the end, researchers aim to give public access to developed AI technology for research purposes.
Within this project we will develop artificial intelligence (AI) methods to improve histological detection of (hereditary) diffuse-type gastric cancer (DGC). Since DGC may be easily missed or hard to find on biopsies and prophylactic hereditary gastrectomy specimens, AI will aid the pathologist in the diagnostic work-up, improving the detection of relevant cell types among a very large set of slides, with high potential to improve cancer diagnostics.
To build these models, we have established an international collaboration to curate a substantial and diverse dataset of images from biopsies and PTGs of (H)DGC patients. This effort involves eight medical centers, seven of which have already contributed data. We anticipate collecting over 800 images from more than 200 patients, making this the largest dataset of (H)DGC patients available for AI research to date.
Recently, we completed our first project, which focused on developing AI models capable of detecting multiple tumor cell types in PTG images. For this, we annotated 350 whole-slide images from 46 patients across four European medical centers, resulting in over 90,000 annotated tumor cells. Additionally, we organized a study in which a panel of pathologists annotate DGC lesions, enabling us to quantifying interpathologist agreement on the tumor cell detection task. When compared with our model’s predictions, we found that the models consistently match or exceed the pathologist-level performance.
In the coming months, we will finish data collection and build models for subsequent projects, which will expand on our initial work by incorporating data from biopsies and advanced-stage cases. These projects aim to develop models that can not only detect DGC lesions in a clinical setting, but also distinguish between sporadic and hereditary DGC based on histology images. Furthermore, we have developed a versatile software platform that integrates with our institute’s diagnostic systems and can automatically execute our AI models, paving the way for deployment in clinical workflows.
Within this project we will develop artificial intelligence (AI) methods to improve histological detection of (hereditary) diffuse-type gastric cancer (DGC). Since DGC may be easily missed or hard to find on biopsies and prophylactic hereditary gastrectomy specimens, AI will aid the pathologist in the diagnostic work-up, improving the detection of relevant cell types among a very large set of slides, with high potential to improve cancer diagnostics.
To develop these AI models, we have collected more than 500 whole-slide images (WSIs) containing DGC tumors from more than 100 different patients. These cases originate from four different European medical centers. Additionally, we are setting up a large international collaboration with nine medical centers to collect a substantial and diverse dataset that includes images of biopsies and PTGs from (H)DGC patients. This valuable dataset will be the basis for later work packages.
More than 300 WSIs have been fully annotated at this point. We have used subsets of this set to train a variety of different tumor cell detection models. One of these was presented at the European Congress of Pathology in 2023. We expect to finish annotation of the full dataset for WP1 in Q1 2024, subsequently enabling us to train and evaluate the complete AI models and finish WP1. In parallel, we organized a reader study to quantify inter-pathologist agreement on the task of multi-class tumor cell detection in patches of HDGC tumors. Five expert pathologists participated in this study, and we are currently analyzing the results.
In the next months, we will finish training the AI models for WP1 and combine it with the results of the reader study to complete this work package. Subsequently, we will start model development for WP2 and continue data collection for WP2 and WP3.
Within this project we will develop artificial intelligence (AI) methods to improve histological detection of (hereditary) diffuse-type gastric cancer (DGC). Since DGC may be easily missed or hard to find on biopsies and prophylactic hereditary gastrectomy specimens, AI will aid the pathologist in the diagnostic work-up, improving the detection of relevant cell types among a very large set of slides, with high potential to improve cancer diagnostics.
To develop these AI models, we have so far collected more than 500 Whole Slide Images (WSIs) containing HDGC tumors from three different Dutch medical centers. We expect to collect more images from other medical centers in the next months.
More than 60 WSIs have been fully annotated at this point. This subset of images is being used to develop an interactive detection framework that will be employed to efficiently annotate the remaining images.
In parallel, we’ve built the pipelines necessary to train and validate the AI models. The pipelines have been made generic, to allow them to be reused in later stages of the project. These pipelines have been tested on smaller subsets of our available data which have shown promising results.
In the next months, we aim to finish annotating a sufficiently large portion of the images and train and evaluate multiple different AI models for the task of signet ring cell detection in Whole Slide Images.