Comprehensive and Clinically Accurate Head and Neck Organs at Risk Delineation via Stratified Deep Learning: A Large-scale Multi-Institutional Study

Accurate organ at risk (OAR) segmentation is critical to reduce the radiotherapy post-treatment complications. Consensus guidelines recommend a set of more than 40 OARs in the head and neck (H&N) region, however, due to the predictable prohibitive labor-cost of this task, most institutions choose a substantially simplified protocol by delineating a smaller subset of OARs and neglecting the dose distributions associated with other OARs. In this work we propose a novel, automated and highly effective stratified OAR segmentation (SOARS) system using deep learning to precisely delineate a comprehensive set of 42 H&N OARs. SOARS stratifies 42 OARs into anchor, mid-level, and small & hard subcategories, with specifically derived neural network architectures for each category by neural architecture search (NAS) principles. We built SOARS models using 176 training patients in an internal institution and independently evaluated on 1327 external patients across six different institutions. It consistently outperformed other state-of-the-art methods by at least 3-5% in Dice score for each institutional evaluation (up to 36% relative error reduction in other metrics). More importantly, extensive multi-user studies evidently demonstrated that 98% of the SOARS predictions need only very minor or no revisions for direct clinical acceptance (saving 90% radiation oncologists workload), and their segmentation and dosimetric accuracy are within or smaller than the inter-user variation. These findings confirmed the strong clinical applicability of SOARS for the OAR delineation process in H&N cancer radiotherapy workflows, with improved efficiency, comprehensiveness, and quality.

Yttrium-90 quantitative phantom study using digital photon counting PET

Background PET imaging of 90Y-microsphere distribution following radioembolisation is challenging due to the count-starved statistics from the low branching ratio of e+/e− pair production during 90Y decay. PET systems using silicon photo-multipliers have shown better 90Y image quality compared to conventional photo-multiplier tubes. The main goal of the present study was to evaluate reconstruction parameters for different phantom configurations and varying listmode acquisition lengths to improve quantitative accuracy in 90Y dosimetry, using digital photon counting PET/CT. Methods Quantitative PET and dosimetry accuracy were evaluated using two uniform cylindrical phantoms specific for PET calibration validation. A third body phantom with a 9:1 hot sphere-to-background ratio was scanned at different activity concentrations of 90Y. Reconstructions were performed using OSEM algorithm with varying parameters. Time-of-flight and point-spread function modellings were included in all reconstructions. Absorbed dose calculations were carried out using voxel S-values convolution and were compared to reference Monte Carlo simulations. Dose-volume histograms and root-mean-square deviations were used to evaluate reconstruction parameter sets. Using listmode data, phantom and patient datasets were rebinned into various lengths of time to assess the influence of count statistics on the calculation of absorbed dose. Comparisons between the local energy deposition method and the absorbed dose calculations were performed. Results Using a 2-mm full width at half maximum post-reconstruction Gaussian filter, the dosimetric accuracy was found to be similar to that found with no filter applied but also reduced noise. Larger filter sizes should not be used. An acquisition length of more than 10 min/bed reduces image noise but has no significant impact in the quantification of phantom or patient data for the digital photon counting PET. 3 iterations with 10 subsets were found suitable for large spheres whereas 1 iteration with 30 subsets could improve dosimetry for smaller spheres. Conclusion The best choice of the combination of iterations and subsets depends on the size of the spheres. However, one should be careful on this choice, depending on the imaging conditions and setup. This study can be useful in this choice for future studies for more accurate 90Y post-dosimetry using a digital photon counting PET/CT.

The Effects of Mismatch between SPECT and CT Images on Quantitative Activity Estimation – A Simulation Study

Background: Quantitative activity estimation is essential in targeted radionuclide therapy dosimetry. Misregistration between SPECT and CT images at the same imaging time point due to patient movement degrades accuracy. This work aims to study the mismatch effects between CT and SPECT data on attenuation correction (AC), volume-of-interest (VOI) delineation and registration for activity estimation.Methods: Nine 4D XCAT phantoms were generated at 1, 24, and 144 hrs post In-111 Zevalin injection, varying in activity distributions, body and organ sizes. Realistic noisy SPECT projections were generated by an analytical projection and reconstructed with quantitative OS-EM method. CT images were shifted from -5 to 5 voxels as well as according to clinical reference corresponding to SPECT images at each time point. For AC effect, mismatched CT images were used for AC in SPECT reconstruction while VOIs were mapped out from matched CTs. For VOI effect, target organs were mapped out using mismatched CTs with matched CTs for AC. For registration effect, non-rigid registrations were performed on sequential mismatched CTs to align corresponding SPECT images, with no AC and VOI mismatch. Bi-exponential curve fitting was performed to obtain time-integrated activity (TIA). Organ activity errors (%OAE) and TIA errors (%TIAE) were calculated.Results: According to clinical reference, %OAE was larger for organs near ribs for AC effect, e.g., -2.58%±0.81% for liver. For VOI effect, %OAE was larger for small and low uptake organs, e.g., -11.94%±10.34% for spleen. %OAE was proportional to mismatch magnitude, e.g., 4.77%±1.41%, 12.01%±3.97% and 42.81%±6.38% for 1-, 2-, and 5-voxel mismatch for lungs. For registration effect, %TIAE were larger when mismatch existed in more numbers of SPECT/CT images, while no substantial difference was observed when using mismatched CT at different time points for registration reference. %TIAE was highest for VOI, followed by registration and AC, e.g., 37.61%±5.08%, 14.25%±7.07% and 1.13%±0.90% respectively for kidneys.Conclusions: The mismatch between CT and SPECT images poses a significant impact on accuracy of quantitative activity estimation in dosimetry, attributed particularly from VOI delineation errors. It is recommended to perform registration between emission and transmission images at the same time point to ensure dosimetric accuracy.