| Creators: |
Hieber, Daniel and Müller, Dominik and Spiess, Ellena and Holl, Felix and Liesche-Starnecker, Friederike and Pryss, Rüdiger and Schobel, Johannes |
| Title: |
Does size and colour matter in computational pathology? First results |
| Item Type: |
Conference or Workshop Item |
| Event Title: |
(Abstracts of the) 37th European Congress of Pathology "Tradition meets Future" |
| Event Location: |
Vienna, Austria |
| Event Dates: |
September, 6-10, 2025 |
| Projects: |
DigiHealth |
| Type of Paper / Paper No.: |
/ Abstract E--PS-08-041 |
| Date: |
2025 |
| Divisions: |
Gesundheitsmanagement |
| Abstract (ENG): |
Background & Objectives: Computational pathology is gaining
increasing popularity, expanding both its clinical applications and computational demands. Unlike radiology, digital pathology relies on extremely high-resolution, colour-rich images, resulting in substantial storage and processing requirements. This raises a critical question: are high magnifcations and colour information truly essential for computational pathology models?
Methods: We analysed three open-access datasets: RINGS (prostate cancer, tile-based, binary segmentation), CoCaHis (colon cancer, tile-based, binary segmentation), and PANDA (prostate cancer, whole-slide images, multi-class segmentation). Each dataset was used to train a
U-Net model implemented in PyTorch Lightning with MONAI, utilizing both colour and grayscale inputs. For PANDA, models were trained at 5x (colour + grayscale) and 10x (only colour) magnifcations. An approximate 80/20 train-test split was applied. To ensure consistency in colour models, H&E staining was normalized using the modifed Reinhard method.
Results: On validation data, the PANDA model trained at 5x outperformed the 10x model (Dice: 68.74% vs. 64.49%), but this trend reversed on the test set (47.10% vs. 63.94%). For CoCaHis, colour-based models consistently outperformed grayscale models (validation Dice: 80.07% vs. 50.87%; test Dice: 51.38% vs. 39.80%). The same pattern was observed for the RINGS dataset (validation Dice: 59.13% vs. 52.36%, test Dice: 61.42% vs 50.96%). For PANDA at 5x magnification, the colour model performed better during validation (68.74% vs. 63.81%) but worse during testing (47.10% vs. 61.95%).
Conclusion: Color input generally improved performance, though not
universally. Similarly, higher resolution enhanced test performance in the PANDA dataset. These fndings align with current best practices but suggest that the benefts from high resolution and colour may not
always be consistent. As these are preliminary results, further investigation is needed to determine when such enhancements are truly justifed. Limitations such as inprecise labels, the use of only three datasets (with only one whole-slide image dataset) should be addressed in following work. |
| Forthcoming: |
No |
| Language: |
English |
| Citation: |
Hieber, Daniel and Müller, Dominik and Spiess, Ellena and Holl, Felix and Liesche-Starnecker, Friederike and Pryss, Rüdiger and Schobel, Johannes
(2025)
Does size and colour matter in computational pathology? First results.
In: (Abstracts of the) 37th European Congress of Pathology "Tradition meets Future", September, 6-10, 2025, Vienna, Austria, Paper / Abstract E--PS-08-041.
(Virchows Archiv).
|
