The present study provides evidence for the substantial benefits of using nnU-Net, a publicly available CNN [9]. A robust positive correlation was demonstrated between manual segmentation and AI segmentation, with limited disparities in the methodologies. While statistically significant differences were observed, they were quantitatively trivial and arose primarily from slight variations in contour definition, as in the case of calcification at the interface of two contours, such as the ascending aorta and aortic arch, leading to contrasting calcium scores based on the contour delineation.

This study has several limitations. The use of masking techniques was necessary due to the pattern of Ga-68 DOTATATE tracer uptake, and manual segmentation remains inherently variable despite efforts to standardize anatomical and non-anatomical contour boundaries. Challenges arise when attempting to differentiate the inner and outer aortic walls due to the thin nature of the vessel lining, potential partial volume effects (leading to potential inaccuracies in quantifying tracer uptake and inflammation) and the use of a CT performed for the purposes of attenuation correction. These limitations may result in small variations in predicted contours of complex structures, leading to significant differences in contouring (as was observed with outliers in aortic and most diseased segment contours) and may have implications for diagnostic, therapeutic planning, and research purposes. Further investigation is necessary to determine the optimal size of the training dataset for clinical use, establish the acceptable level of variability, and clarify the role of the clinician in advanced segmentation (especially in vascular territories), considering the presence of intravascular stents, surgical material, and variant or rare anatomy.

Small vascular lesions are prone to partial volume artifact, low target-to-background ratios, and the proximity of the blood pool which represent challenges to tracer development and image interpretation, constraints gradually being addressed by technological advances including total body PET, new image reconstruction and motion correction techniques and hybrid tracer imaging using nanoparticles [10,11,12].

Results of the present work align with previous research reporting significant improvements in both analysis time and contour accuracy with the use of a CNN compared to manual segmentation [13]. The performance of nnU-Net (in particular) has been extensively evaluated for non-vascular segmentation and has consistently demonstrated strong performance across a wide range of applications, including neuroradiology, cardiology, musculoskeletal injury and oncology [14,15,16,17]. Manual segmentation is known to be a labour-intensive process and the present study demonstrates the significant decrease in workflow time that can be achieved. Deep learning based automated segmentation has the potential to improve efficiency and reproducibility to a clinically acceptable standard equal to, or even greater, than can be attained by trained clinicians [18, 19]. Implementing such techniques in the clinical workflow would both streamline and improve quality of diagnosis, utilisation of senior clinician time and improve the access and affordability of non-invasive functional imaging. This reduction in time represents an improvement in the utilization of a skilled workforce, increasing efficiency, and enhancing productivity. Furthermore, these findings translate to cost savings for healthcare services. Importantly, the performance of the model was not significantly impaired across a diverse patient population, including those with and without severe vascular calcification.

One of the most cited, and foundational CNNs used for deep-learning experimentation in medical imaging is the U-Net architecture [20]. The network is comprised of a contracting and expanding path, symmetric in their use of down- and up-sampling operators, giving the model it’s identifiable "U" shape. U-Net and its variants have demonstrated high accuracy in segmenting biomedical images and wide applicability [21,22,23,24,25,26,27]. Rapid advancement across a range of both open-source and proprietary AI models has led to advances in CT-FFR, improvements in cardiovascular event prediction by nuclear perfusion imaging, and myocardial tissue characterisation on cardiac MRI [28, 29]. However, replicating published benchmarks requires careful modification of model configurations and training schemes, catering to the characteristics of choice datasets. This is especially prevalent in three-dimensional biomedical imaging problem domains, where imaging modality, anisotropic voxel spacing, and imaging dimensionality may vary dramatically between facilities. The high dimensionality of model hyperparameter configuration, coupled with the limited supply of training and validation data, thus often leads to models failing to live up to their promised performance when evaluated on similar, but unseen problem domains.

nnU-Net aids in navigating this complex parameter domain by handling all the pre-processing, training, and inference-making in the prediction pipeline. To achieve this, design choices are based on the data itself considering such factors as voxel-spacing, image dimensions and class ratios. Streamlining this greatly reduces the hyperparameter tuning domain, allowing the user to quickly generate literature-comparable results and build from there.

The use of AI in medical imaging represents a major stride forward in the advancement of healthcare. The present study serves as a demonstration of the capabilities of AI and its ability to provide more efficient and accurate results than traditional methods. The findings of this study have significant implications for the field of medical imaging and provide compelling evidence for the continued investment in research and development of AI in the healthcare sector. With the continued advancement of AI technologies, the ability to analyse complex medical images and generate data from large datasets will only continue to improve. This, in turn could streamline clinical trial conduct, provide a platform for personalised medicine and ultimately, improve health outcomes for patients. New knowledge gained from this study includes demonstration of strong, positive agreement in clinical measures of vascular tracer uptake and calcification beyond contour Dice coefficients, alongside a significant reduction in clinician workload.