Abstract
Introduction:
Deep learning models benefit from large and varied, but costly, datasets. This study investigates the dataset size trade-off in the context of pelvic multi-organ MR segmentation; specifically assessing the performance of a well-known segmentation model, nnU-Net, in a limited domain and limited dataset setting.
Materials/methods:
12 participants undergoing treatment on an Elekta Unity were recruited, acquiring 58 MR images, with 4 participants (12 images) withheld for testing. Prostate, seminal vesicles (SV), bladder and rectum were contoured in each image by a radiation oncologist. Six more models were trained on progressively smaller subsets of the training dataset, simulating a restricted dataset setting. To observe the efficacy of data augmentation, another set of identical models were trained without augmentation. The performance of the networks was evaluated via the Dice Similarity Coefficient, mean surface distance and 95% Hausdorff distance metrics.
Results:
When trained with entire training dataset (46 images), the model achieved a mean Dice coefficient of 0.903 (Prostate), 0.851 (SV), 0.884 (Rectum) and 0.967 (Bladder). Segmentation performance remained stable when the number of training sets was >12 images from 4 participants, but rapidly dropped in smaller data subsets. Data augmentation was found to be influential across all dataset sizes, but especially in very small datasets.
Summary:
This study demonstrated nnU-Net's ability in performing pelvic multi-organ segmentation in a limited domain and limited data context. We conclude that while our model may not generalise well to scanner or protocol changes, the protocol’s low data requirement can be advantageous for in-house cases with a consistently narrow domain. For example, it could assist in treatment planning by initiating the contouring process instead of manually starting from the beginning.