Synthetic cranial MRI from 3D optical surface scans using deep learning for radiation therapy treatment planning

• DOI: 10.1007/s13246-023-01229-4
• Publication Date: 2023-02-08T16:45:48Z
• Authors: Michael Douglass; Peter Gorayski; Sandy Patel; Alexandre Santos
• Keywords: Ground truth

Optical scanning technologies are increasingly being utilised to supplement treatment workflows in radiation oncology, such as surface-guided radiotherapy or 3D printing custom bolus. One limitation of optical devices is the absence internal anatomical information patient scanned. As a result, conventional therapy planning using this imaging modality not feasible. Deep learning useful for automating various manual tasks most notably, organ segmentation and planning. models have also been used transform MRI datasets into synthetic CT datasets, facilitating development MRI-only planning.To train pix2pix generative adversarial network scan data estimated given provide additional select few sites. The proposed may surface mould brachytherapy, total body irradiation, skin electron therapy, example, without delivering any dose.A 2D GAN was trained on 15,000 axial slices healthy adult brains paired with corresponding external mask slices. model validated further 5000 previously unseen predictions were compared "ground-truth" multi-scale structural similarity index (MSSI) metric. A certified neuro-radiologist subsequently consulted an independent review model's performance terms accuracy consistency. then applied photogrammetry test subject demonstrate feasibility novel technique.The predicted mean MSSI 0.831 ± 0.057 validation images indicating that it possible estimate significant proportion patient's gross cranial anatomy from exterior contour. When independently reviewed by neuro-radiologist, described "quite amazing, but there limitations regions where wide variation within normal population." human acquired photogrammetry, could volume good qualitative accuracy. However, ground-truth baseline available quantitative comparison.A deep developed, volume, potentially increasing usefulness This work has demonstrated much can be shape head source valuable data. Further research required investigate approach use clinical setting improve