Machine learning to predict environmental dose rates from a radionuclide therapy service - a proof of concept study.

The Ionising Radiation Regulations 2017 requires prior risk assessment calculations and regular environmental monitoring of radiation doses. However, the accuracy of prior risk assessments is limited by assumptions and monitoring only provides retrospective evaluation. This is particularly challenging in nuclear medicine for areas surrounding radionuclide therapy patient bathroom wastewater pipework. Machine learning (ML) is a technique that could be applied to patient booking records to predict environmental radiation dose rates in these areas to aid prospective risk assessment calculations, which this proof-of-concept work investigates. 540 days of a dosimeters historical daily average dose rate measurements and the corresponding period of department therapy booking records were used to train six different ML models. Predicted versus measured daily average dose rates for the following 60 days were analysed to assess and compare model performance. A wide range in prediction errors was observed across models. The gradient boosting regressor produced the best accuracy (root mean squared error = 1.10µSv.hr-1, mean absolute error = 0.87µSv.hr-1, mean absolute percentage error = 35% and maximum error = 3.26µSv.hr-1) and goodness of fit (R2= 0.411). Methods to improve model performance and other scenarios where this approach could prove more accurate were identified. This work demonstrates that ML can predict temporal fluctuations in environmental radiation dose rates in the areas surrounding radionuclide therapy wastewater pipework and indicates that it has the potential to play a role in improving legislative compliance, the accuracy of radiation safety and use of staff time and resources.

Patient-specific optimisation of administered activity and acquisition times for 18F-FDG PET imaging.

BACKGROUND: The purpose of this study is to identify a method for optimising the administered activity and acquisition time for 18F-FDG PET imaging, yielding images of consistent quality for patients with varying body sizes and compositions, while limiting radiation doses to patients and staff. Patients referred for FDG scans had bioimpedance measurements. They were injected with 3 MBq/kg of 18F up to 370 MBq and scanned on a Siemens Biograph mCT at 3 or 4 min per bed position. Data were rebinned to simulate 2- and 1-min acquisitions. Subjective assessments of image quality made by an experienced physician were compared with objective measurements based on signal-to-noise ratio and noise equivalent counts (NEC). A target objective measure of image quality was identified. The activity and acquisition time required to achieve this were calculated for each subject. Multiple regression analysis was used to identify expressions for the activity and acquisition time required in terms of easily measurable patient characteristics. RESULTS: One hundred and eleven patients were recruited, and subjective and objective assessments of image quality were compared for 321 full and reduced time scans. NEC-per-metre was identified as the objective measure which best correlated with the subjective assessment (Spearman rank correlation coefficient 0.77) and the best discriminator for images with a subjective assessment of "definitely adequate" (area under the ROC curve 0.94). A target of 37 Mcount/m was identified. Expressions were identified in terms of patient sex, height and weight for the activity and acquisition time required to achieve this target. Including measurements of body composition in these expressions was not useful. Using these expressions would reduce the mean activity administered to this patient group by 66 MBq compared to the current protocol. CONCLUSIONS: Expressions have been identified for the activity and acquisition times required to achieve consistent image quality in FDG imaging with reduced patient and staff doses. These expressions might need to be adapted for other systems and reconstruction protocols.

Evaluation of image enhancement software as a method of performing half-count bone scans.

Reducing the radiation dose and scanning time of diagnostic tests is often desirable. One method uses image enhancement software such as Pixon, which processes lower-count scans and aims to produce high-quality images. However, it is essential that diagnostic accuracy is not compromised. We compared the level of agreement between clinicians using standard scans, with half-count and Pixon-enhanced half-count scans. Bone scans from 150 patients referred to diagnose metastatic disease were degraded by a process of Poisson-preserving binomial resampling to generate equivalent half-count scans and then processed by Pixon software to recreate 'original' high-quality scans. Two experienced clinicians reported the scans in a randomized, blinded manner for metastatic disease (yes/no) and assigned a confidence level to this diagnosis. Levels of agreement between clinicians were calculated for the full-count, half-count, and Pixon-enhanced half-count scans and between scanning methods for each clinician. Agreement between clinicians for standard full-count scans was 92% (±4%, κ=0.80), compared with 92% (±4%, κ=0.79) for half-count scans and 87% (±5%, κ=0.70) for Pixon-processed half-count scans. Agreement for a single clinician viewing full-count versus half-count scans was 95% (±2%, κ=0.88), similar to the agreement for a single clinician viewing full-count versus Pixon-processed half-count scans (95%, ±2%, κ=0.88). With respect to confidence in diagnosis, 127 full-count scans were scored in the highest category, compared with 98 half-count and 88 Pixon-processed half-count scans. Switching to half-count scanning does not introduce more diagnostic disagreement than is already present between clinicians. However, clinicians feel less confident reporting half-count scans. The Pixon enhancement step improved neither objective diagnostic agreement nor clinician confidence.