The feasibility of training nnU-Net to segment pediatric Wilms tumor with small datasets.
Authors
Affiliations (4)
Affiliations (4)
- Princess Máxima Center for Pediatric Oncology, Utrecht, Netherlands.
- Division of Childhealth, Wilhelmina Childrens' Hospital, Utrecht, Netherlands.
- Department of Radiology and Nuclear Medicine, University Medical Center Utrecht, Utrecht, Netherlands.
- Image Sciences Institute, University Medical Center Utrecht, Utrecht University, Utrecht, Netherlands.
Abstract
Segmentation of magnetic resonance imaging (MRI) can be used in the clinical care of patients with (suspected) Wilms tumor (WT). Manual segmentation, however, is time consuming and user dependent. Deep learning can be used to automate this, but is often associated with the need for large training datasets to be trained successfully, which is inherently challenging in pediatric populations. The aim of this study was to determine the effect of the number of training patients on automated WT segmentation. 56 patients with a suspected WT were included in this retrospective study. The pre-operative MRI scans were divided in a training (n = 45) and test (n = 11) set. The segmentation was based on nnU-Net, an out-of-the-box framework for image segmentation. nnU-Net was trained on a random selection of five patients, and the resulting segmentation method was tested on a fixed test set of 11 patients. Next, patients were added to the training dataset with incremental steps of five patients, up to the full dataset of 45 patients. This process was repeated 5 times to account for random sampling. The Dice score was used as measure for segmentation performance. We pragmatically defined a Dice score of > 0.80 as a reference level for segmentation performance. The lowest median Dice score was obtained when using a subset of five training patients (0.38). The highest median Dice score was observed for training on 45 patients (0.91). An median Dice score of 0.80 was reached using a minimal subset of 10 training patients. The Dice score showed dependency on the random selection of the training patients, with the largest spread in performance observed in the training set of five patients. A large spread in performance between the 11 test set patients was observed. Using this specific training and test cohort, a median Dice score of 0.80 was reached with just 10 training patients, however, the results showed considerable variation in the individual test patients. The performance spread between the random subsampling of the training patients highlights the importance of clinically representable datasets when dealing with the small dataset sizes inherent to pediatric oncology.