| Creators: |
Hieber, Daniel and Prokop, Georg and Karthan, Maximilian and Märkl, Bruno and Schobel, Johannes and Liesche-Starnecker, Friederike |
| Title: |
Neural Network Assisted Pathology for Labeling Tumors in Whole-Slide-Images of Glioblastoma |
| Item Type: |
Conference or Workshop Item |
| Event Title: |
(Abstractband der) 106. Jahrestagung der Deutschen Gesellschaft für Pathologie "Pathology - more than meets the eye" |
| Event Location: |
Leipzig, Germany |
| Event Dates: |
1.-3. Juni 2023 |
| Projects: |
NAP, DigiHealth |
| Page Range: |
p. 218 |
| Additional Information: |
Abstract P.13.13 |
| Date: |
2023 |
| Divisions: |
Gesundheitsmanagement |
| Abstract (ENG): |
Introduction:
Most segmentation models that are based on Machine Learning (ML) focus on lowerresolution
MRI images rather than high-resolution histopathological images, especially for
brain tumors [1]. The main reason for this is the reduced labeling effort required to train
models for lower-resolution imaging.
Objective:
To utilize the large data available in pathology, we trained an ML model to handle the
segmentation of glioblastoma in hematoxylin and eosin (HE) slides. The trained model is then
capable to conduct labeling of histopathological images without the need for manual
annotation by domain experts.
Methods:
We used 103 hematoxylin and eosin-stained high-resolution Whole-Slide-Images (WSI) of 56
patients whose tumor-containing regions had been annotated by a neuropathologist. The
images were tiled into processable chunks of 1024x1024 pixels and rescaled to 512x512
pixels. We used the „Segmentation Models“-framework [2], a simple-to-use tool for image
segmentation and classification in PyTorch [3], to conduct the segmentation.
Results:
The tumor tiles were segmented with an Intersection-over-Union (IoU) of 84.5% after only 7
training epochs. Better preprocessing algorithms (e.g., removing small errors in the labels)
can further increase the IoU and are scope of future research. Moreover, a generic
segmentation model was used, that was not specifically tailored to our application scenario.
With specific fine-tuning, the accuracy may also be increased.
Conclusion:
Based on a dataset of labeled, HE-stained WSIs, we could segment these into tumorcontaining
and non-neoplastic tissue with an IoU of 84.5% using a preconfigured ML model.
Utilizing this initial manual labeling of 103 WSIs, future labeling tasks can be handled
automatically via ML. While the current segmentation score is not sufficient for daily routine
diagnostic, it can already be used for preprocessing tasks (e.g., selecting relevant tissue for
future analysis). Adjustments in the applied algorithms will increase the accuracy, allowing a
broader use.
The prototype shows the enormous potential in histopathological data, allowing accurate
analysis with ML models with rather small effort. While the presented model was trained
using glioblastoma WSIs, the applicability for other tumors is easily achievable with a few
dozen labeled WSIs. |
| Forthcoming: |
No |
| Language: |
English |
| Link eMedia: |
Download |
| Citation: |
Hieber, Daniel and Prokop, Georg and Karthan, Maximilian and Märkl, Bruno and Schobel, Johannes and Liesche-Starnecker, Friederike
(2023)
Neural Network Assisted Pathology for Labeling Tumors in Whole-Slide-Images of Glioblastoma.
In: (Abstractband der) 106. Jahrestagung der Deutschen Gesellschaft für Pathologie "Pathology - more than meets the eye", 1.-3. Juni 2023, Leipzig, Germany, p. 218.
|
