Session Item

High-resolution (HR) MRI provides detailed soft-tissue information that is useful in assessing long-term side-effects after radiotherapy in childhood cancer survivors, such as facial asymmetry or morphological changes in brain structures. However, 3D HRMRI requires long acquisition times, so in practice often multiple 2D low-resolution (LR) images (with thick slices in multiple planes) are acquired for patient follow-up.

In this work, we present a super-resolution (SR) convolutional neural network (CNN) which can reconstruct a HR 3D image from 2D LR images, in order to improve the extraction of structural biomarkers from routine scans.

A multi-level densely connected super-resolution CNN [1] was adapted to take two perpendicular LR scans (e.g., coronal and axial) as tensors and reconstruct a 3D HR image. Scans were resampled to a resolution of 1mm³ before being fed into the network. A training set of 80 HR T1 paediatric (9-10years, healthy subjects) head scans from the Adolescent Brain Cognitive Development (ABCD) study was used as baseline, and 2D LR images were simulated to use as input into the CNN (Figure 1). 10 additional scans from ABCD were used to tune the hyperparameters of the CNN. The output of the model (images_CNN) was compared against simple interpolation (resampling and averaging both inputs), (images_interp).

The evaluation was done in two steps. First, the quality of the reconstructed HR images was assessed using the peak signal-to-noise ratio (PSNR) (larger values indicate better quality) compared to baseline. Secondly, the precision of structure segmentation (using the autocontouring software Limbus AI) in the reconstructed vs the baseline HR images was assessed using mean distance-to-agreement (mDTA). As Limbus AI is not validated for paediatric data, we carefully inspected the segmentations.

Three datasets were available: 1) 10 new ABCD images (dataset 1); 2) 18 images from the Children’s Brain Tumour Network (CBTN) study (acquired HR and simulated LR images, age 2–20years, dataset 2) and 3) 6 “real-world” follow-up images of a paediatric head and neck cancer patient (acquired HR and acquired LR, 14-19years, dataset 3).

The proposed CNN outperformed simple interpolation. PSNR for images_CNN were on average(sd) 26.1(2.1) for dataset 1 and 24.4(2.6) for dataset 2, while for all images_interp were 20.5(1.9) and 21.4(2.8), respectively.

Similarly, structure segmentation was more precise (closer to that of baseline images) in images_CNN compared to images_interp (Figures 2a and 2b).

This work demonstrates that deep learning methods can successfully reconstruct 3D HR images from 2D LR ones, potentially advancing research for paediatric radiotherapy effects. Our model outperforms standard interpolation, both in perceptual quality and for autocontouring. Further work is needed to improve the generalisability across imaging sequences and validate it for additional structural analysis tasks.

[1] https://arxiv.org/abs/1803.01417