Kamp, A. van der (2021) Deep learning-based histopathologic assessment of Wilms Tumors. thesis, Medicine.
Full text available on request.Abstract
Background: Wilms Tumor (WT) or nephroblastoma is the most common type of renal cancer in children. Histopathological assessment is crucial for risk group classification to determine postoperative treatment regimens. However, research has shown a discrepancy between pathologists in the quantification of WT components. The interobserver variability in the quantification of WT can lead to a misclassification and therefore compromise the optimal treatment. As a result, patients could be at a higher risk of disease recurrence (undertreatment) or at an increased risk of toxic effects later on (overtreatment). Therefore, the quantification of WT components could highly benefit from an accurate and uniform assessment. Artificial Intelligence has been shown in recent studies to contribute to a more accurate and improved classification in assessments that are highly subject to interobserver variability. Objective: To assess the performance of a deep learning (DL)-based system in quantifying WT in haematoxylin and eosin-stained tissue. Methods: In collaboration with Radboud University Medical Center an algorithm to quantify WT-elements has been developed. The annotations, in a cohort of 74 patients, were done by trained medical students and were checked by an expert pathologist after the interobserver variability was determined (κ ³ 0.81). The dataset contained 4995 annotations in total, divided between nineteen tissue elements (four tissue elements were not investigated further due to infrequent occurrence). The database was divided in 70% training (to train the algorithm), 15% validation (to fine tune the algorithm) and 15% testing (to evaluate the performance of the algorithm). The accuracy of the DL-based system was calculated with the Sørensen-Dicecoefficient. The performance of the algorithm was determined by the number of correct and incorrect predictions, which was displayed in a confusion matrix. Results: The overall accuracy of the DL-based system in distinguishing fifteen tissue components was 85%. The overall accuracy of the vital tumor components is 70% and for chemotherapy induced changes 84%. The DL-based system yielded F1 scores among the fifteen tissue components ranging from 0.36 (bleeding) – 0.99 (lymph nodes). The discordance in the risk-group between (expert) pathologists and the DL-based system is 13%. Conclusion: WT and normal kidney tissue components can be automatically recognized by DL with high accuracy. However, there is still a significant discordance between the conventional assessment of the pathologists and the automatic quantification. In addition, before the automatic quantification can be used both methods should be correlated to patient outcome.
| Item Type: | Thesis (UNSPECIFIED) |
|---|---|
| Supervisor name: | Kuks, Prof. Dr. J.B.M. and Krijger, Dr. R.R. de and Mavinkurve-Groothuis, Dr. A.M.C. |
| Faculty: | Medical Sciences |
| Date Deposited: | 24 Dec 2021 10:26 |
| Last Modified: | 24 Dec 2021 10:26 |
| URI: | https://umcg.studenttheses.ub.rug.nl/id/eprint/2908 |
Actions (login required)
 |
View Item |