Deep learning for contour quality assurance for RTOG 0933: In-silico evaluation.

PURPOSE: To validate a CT-based deep learning (DL) hippocampal segmentation model trained on a single-institutional dataset and explore its utility for multi-institutional contour quality assurance (QA). METHODS: A DL model was trained to contour hippocampi from a dataset generated by an institutional observer (IO) contouring on brain MRIs from a single-institution cohort. The model was then evaluated on the RTOG 0933 dataset by comparing the treating physician (TP) contours to blinded IO and DL contours using Dice and Haussdorf distance (HD) agreement metrics as well as evaluating differences in dose to hippocampi when TP vs. IO vs. DL contours are used for planning. The specificity and sensitivity of the DL model to capture planning discrepancies was quantified using criteria of HD > 7 mm and Dmax hippocampi > 17 Gy. RESULTS: The DL model showed greater agreement with IO contours compared to TP contours (DL:IO L/R Dice 74 %/73 %, HD 4.86/4.74; DL:TP L/R Dice 62 %/65 %, HD 7.23/6.94, all p < 0.001). Thirty percent of contours and 53 % of dose plans failed QA. The DL model achieved an AUC L/R 0.80/0.79 on the contour QA task via Haussdorff comparison and AUC of 0.91 via Dmax comparison. The false negative rate was 17.2 %/20.5 % (contours) and 5.8 % (dose). False negative cases tended to demonstrate a higher DL:IO Dice agreement (L/R p = 0.42/0.03) and better qualitative visual agreement compared with true positive cases. CONCLUSION: Our study demonstrates the feasibility of using a single-institutional DL model to perform contour QA on a multi-institutional trial for the task of hippocampal segmentation.

Synthetic pulmonary perfusion images from 4DCT for functional avoidance using deep learning.

Purpose.To develop and evaluate the performance of a deep learning model to generate synthetic pulmonary perfusion images from clinical 4DCT images for patients undergoing radiotherapy for lung cancer.Methods. A clinical data set of 58 pre- and post-radiotherapy99mTc-labeled MAA-SPECT perfusion studies (32 patients) each with contemporaneous 4DCT studies was collected. Using the inhale and exhale phases of the 4DCT, a 3D-residual network was trained to create synthetic perfusion images utilizing the MAA-SPECT as ground truth. The training process was repeated for a 50-imaging study, five-fold validation with twenty model instances trained per fold. The highest performing model instance from each fold was selected for inference upon the eight-study test set. A manual lung segmentation was used to compute correlation metrics constrained to the voxels within the lungs. From the pre-treatment test cases (N = 5), 50th percentile contours of well-perfused lung were generated from both the clinical and synthetic perfusion images and the agreement was quantified.Results. Across the hold-out test set, our deep learning model predicted perfusion with a Spearman correlation coefficient of 0.70 (IQR: 0.61-0.76) and a Pearson correlation coefficient of 0.66 (IQR: 0.49-0.73). The agreement of the functional avoidance contour pairs was Dice of 0.803 (IQR: 0.750-0.810) and average surface distance of 5.92 mm (IQR: 5.68-7.55).Conclusion. We demonstrate that from 4DCT alone, a deep learning model can generate synthetic perfusion images with potential application in functional avoidance treatment planning.

Hippocampus segmentation on noncontrast CT using deep learning.

PURPOSE: Accurate segmentation of the hippocampus for hippocampal avoidance whole-brain radiotherapy currently requires high-resolution magnetic resonance imaging (MRI) in addition to neuroanatomic expertise for manual segmentation. Removing the need for MR images to identify the hippocampus would reduce planning complexity, the need for a treatment planning MR imaging session, potential uncertainties associated with MRI-computed tomography (CT) image registration, and cost. Three-dimensional (3D) deep convolutional network models have the potential to automate hippocampal segmentation. In this study, we investigate the accuracy and reliability of hippocampal segmentation by automated deep learning models from CT alone and compare the accuracy to experts using MRI fusion. METHODS: Retrospectively, 390 Gamma Knife patients with high-resolution CT and MR images were collected. Following the RTOG 0933 guidelines, images were rigidly fused, and a neuroanatomic expert contoured the hippocampus on the MR, then transferred the contours to CT. Using a calculated cranial centroid, the image volumes were cropped to 200 × 200 × 35 voxels, which were used to train four models, including our proposed Attention-Gated 3D ResNet (AG-3D ResNet). These models were then compared with results from a nested tenfold validation. From the predicted test set volumes, we calculated the 100% Hausdorff distance (HD). Acceptability was assessed using the RTOG 0933 protocol criteria, and contours were considered passing with HD ≤ 7 mm. RESULTS: The bilateral hippocampus passing rate across all 90 models trained in the nested cross-fold validation was 80.2% for AG-3D ResNet, which performs with a comparable pass rate (P = 0.3345) to physicians during centralized review for the RTOG 0933 Phase II clinical trial. CONCLUSIONS: Our proposed AG-3D ResNet's segmentation of the hippocampus from noncontrast CT images alone are comparable to those obtained by participating physicians from the RTOG 0933 Phase II clinical trial.

Analysis of homogeneous layer approximations in the modeling of localized surface plasmon resonance biosensors.

We analyze theoretically the effectiveness of homogeneous layer approximations in modeling localized surface plasmon resonance biosensors made of spherical metal nanoparticles coated with biomolecular layers that have radially variable refractive indices. Using an extended Mie theory, we compute the extinction spectrum and peak wavelength of the system and compare them with when effective medium approximations are applied to treat the biomolecular layer as homogeneous. We investigate how the accuracies of the approximations depend on the geometric parameters of the system and the material of the metal nanoparticle. We also derive expressions that can be used to predict if the approximations would accurately predict the spectral position of the peak wavelength.

LSPR Biosensing Approach for the Detection of Microtubule Nucleation.

Microtubules are dynamic protein filaments that are involved in a number of cellular processes. Here, we report the development of a novel localized surface plasmon resonance (LSPR) biosensing approach for investigating one aspect of microtubule dynamics that is not well understood, namely, nucleation. Using a modified Mie theory with radially variable refractive index, we construct a theoretical model to describe the optical response of gold nanoparticles when microtubules form around them. The model predicts that the extinction maximum wavelength is sensitive to a change in the local refractive index induced by microtubule nucleation within a few tens of nanometers from the nanoparticle surface, but insensitive to a change in the refractive index outside this region caused by microtubule elongation. As a proof of concept to demonstrate that LSPR can be used for detecting microtubule nucleation experimentally, we induce spontaneous microtubule formation around gold nanoparticles by immobilizing tubulin subunits on the nanoparticles. We find that, consistent with the theoretical model, there is a redshift in the extinction maximum wavelength upon the formation of short microtubules around the nanoparticles, but no significant change in maximum wavelength when the microtubules are elongated. We also perform kinetic experiments and demonstrate that the maximum wavelength is sensitive to the microtubule nuclei assembly even when microtubules are too small to be detected from an optical density measurement.