Session Item

Paediatric datasets are not readily available for training neural networks for auto-segmentation of organs. Paediatric CT data is also highly variable, not only in patient and anatomy, but in scan dimensions, image quality and acquisition protocols. As such, the performance of neural network algorithms such as U-NET for paediatric images is not yet known, despite being well characterised for adult images. This study aims to (i) evaluate the performance of a segmentation model trained on adult data when applied to paediatric images, (ii) quantify any improvement gained by the inclusion of paediatric data in the training images, and (iii) determine whether variations between CT scanners influence the results.

A total of 1000 adult and 400 Paediatric CT scans with organ contours approved by expert physicians were used for training, validating, and testing the segmentation model: a self-configuring deep learning neural network known as nnU-NET. For each organ/structure, three separate models were generated: one trained on adult data only (A-trained), one on paediatric data only (P-trained) and one trained on both adult and paediatric data (A&P-trained). Paediatric images (<18 years) were acquired on two CT scanners having different acquisition protocols (e.g. helical pitch) and reconstruction algorithms (ASIR vs SAFIRE). A model trained on one scanner set was tested on the images from the other. Test data for 80 adults and 50 paediatrics were withheld from training for evaluation purposes by comparison of Dice Similarity Coefficients relative to expert contours for a selection of head and neck, abdominal, and thoracic organs.

The A-trained model performed poorer on paediatric test data (mean DSC 0.71 – 0.84) than on adult test data (0.83 – 0.93). See Figure 1 for a sample of 7 abdominal and thoracic organs. Poorer performance on the paediatric data is statistically significant (p <0.001). Inclusion of paediatric training data improved this to mean DSCs of 0.79 – 0.97 for the combined A&P-trained model applied to paediatric test data, but the paediatric-specific P-trained model did not have a statistically significant advantage over the A&P-trained model (p=0.265), indicating that there is no need to maintain separate segmentation models for this nn-UNet architecture (see Fig 2). The inter-scanner comparison showed equivalent performance for the two scanner image sets regardless of inclusion in the training data.

The performance of the nnU-NET auto-segmentation model trained only on adult images was satisfactory for application to paediatric CT images (mean DSC > 0.71), however, the inclusion of paediatric images in the training data significantly improved the accuracy of paediatric contouring. The combined-training model performed as well as a paediatric-specific model, indicating that separate adult and paediatric models are not required.