Transfer learning for auto-segmentation of 17 organs-at-risk in the head and neck: Bridging the gap between institutional and public datasets.

BACKGROUND: Auto-segmentation of organs-at-risk (OARs) in the head and neck (HN) on computed tomography (CT) images is a time-consuming component of the radiation therapy pipeline that suffers from inter-observer variability. Deep learning (DL) has shown state-of-the-art results in CT auto-segmentation, with larger and more diverse datasets showing better segmentation performance. Institutional CT auto-segmentation datasets have been small historically (n < 50) due to the time required for manual curation of images and anatomical labels. Recently, large public CT auto-segmentation datasets (n > 1000 aggregated) have become available through online repositories such as The Cancer Imaging Archive. Transfer learning is a technique applied when training samples are scarce, but a large dataset from a closely related domain is available. PURPOSE: The purpose of this study was to investigate whether a large public dataset could be used in place of an institutional dataset (n > 500), or to augment performance via transfer learning, when building HN OAR auto-segmentation models for institutional use. METHODS: Auto-segmentation models were trained on a large public dataset (public models) and a smaller institutional dataset (institutional models). The public models were fine-tuned on the institutional dataset using transfer learning (transfer models). We assessed both public model generalizability and transfer model performance by comparison with institutional models. Additionally, the effect of institutional dataset size on both transfer and institutional models was investigated. All DL models used a high-resolution, two-stage architecture based on the popular 3D U-Net. Model performance was evaluated using five geometric measures: the dice similarity coefficient (DSC), surface DSC, 95th percentile Hausdorff distance, mean surface distance (MSD), and added path length. RESULTS: For a small subset of OARs (left/right optic nerve, spinal cord, left submandibular), the public models performed significantly better (p < 0.05) than, or showed no significant difference to, the institutional models under most of the metrics examined. For the remaining OARs, the public models were inferior to the institutional models, although performance differences were small (DSC ≤ 0.03, MSD < 0.5 mm) for seven OARs (brainstem, left/right lens, left/right parotid, mandible, right submandibular). The transfer models performed significantly better than the institutional models for seven OARs (brainstem, right lens, left/right optic nerve, left/right parotid, spinal cord) with a small margin of improvement (DSC ≤ 0.02, MSD < 0.4 mm). When numbers of institutional training samples were limited, public and transfer models outperformed the institutional models for most OARs (brainstem, left/right lens, left/right optic nerve, left/right parotid, spinal cord, and left/right submandibular). CONCLUSION: Training auto-segmentation models with public data alone was suitable for a small number of OARs. Using only public data incurred a small performance deficit for most other OARs, when compared with institutional data alone, but may be preferable over time-consuming curation of a large institutional dataset. When a large institutional dataset was available, transfer learning with models pretrained on a large public dataset provided a modest performance improvement for several OARs. When numbers of institutional samples were limited, using the public dataset alone, or as a pretrained model, was beneficial for most OARs.

Magnetic resonance biomarker assessment software (MR-BIAS): an automated open-source tool for the ISMRM/NIST system phantom.

Objective.To provide an open-source software for repeatable and efficient quantification ofT1andT2relaxation times with the ISMRM/NIST system phantom. Quantitative magnetic resonance imaging (qMRI) biomarkers have the potential to improve disease detection, staging and monitoring of treatment response. Reference objects, such as the system phantom, play a major role in translating qMRI methods into the clinic. The currently available open-source software for ISMRM/NIST system phantom analysis, Phantom Viewer (PV), includes manual steps that are subject to variability.Approach.We developed the Magnetic Resonance BIomarker Assessment Software (MR-BIAS) to automatically extract system phantom relaxation times. The inter-observer variability (IOV) and time efficiency of MR-BIAS and PV was observed in six volunteers analysing three phantom datasets. The IOV was measured with the coefficient of variation (CV) of percent bias (%bias) inT1andT2with respect to NMR reference values. The accuracy of MR-BIAS was compared to a custom script from a published study of twelve phantom datasets. This included comparison of overall bias and %bias for variable inversion recovery (T1VIR), variable flip angle (T1VFA) and multiple spin-echo (T2MSE) relaxation models.Main results.MR-BIAS had a lower mean CV withT1VIR(0.03%) andT2MSE(0.05%) in comparison to PV withT1VIR(1.28%) andT2MSE(4.55%). The mean analysis duration was 9.7 times faster for MR-BIAS (0.8 min) than PV (7.6 min). There was no statistically significant difference in the overall bias, or the %bias for the majority of ROIs, as calculated by MR-BIAS or the custom script for all models.Significance.MR-BIAS has demonstrated repeatable and efficient analysis of the ISMRM/NIST system phantom, with comparable accuracy to previous studies. The software is freely available to the MRI community, providing a framework to automate required analysis tasks, with the flexibility to explore open questions and accelerate biomarker research.

Comparison of dual-energy CT with positron emission tomography for lung perfusion imaging in patients with non-small cell lung cancer.

Objective.Functional lung avoidance (FLA) radiotherapy treatment aims to spare lung regions identified as functional from imaging. Perfusion contributes to lung function and can be measured from the determination of pulmonary blood volume (PBV). An advantageous alternative to the current determination of PBV from positron emission tomography (PET) may be from dual energy CT (DECT), due to shorter examination time and widespread availability. This study aims to determine the correlation between PBV determined from DECT and PET in the context of FLA radiotherapy.Approach.DECT and PET acquisitions at baseline of patients enrolled in the HI-FIVE clinical trial (ID: NCT03569072) were reviewed. Determination of PBV from PET imaging (PBVPET), from DECT imaging generated from a commercial software (Syngo.via, Siemens Healthineers, Forchheim, Germany) with its lowest (PBVsyngoR=1) and highest (PBVsyngoR=10) smoothing level parameter value (R), and from a two-material decomposition (TMD) method (PBVTMDL) with variable median filter kernel size (L) were compared. Deformable image registration between DECT images and the CT component of the PET/CT was applied to PBV maps before resampling to the PET resolution. The Spearman correlation coefficient (rs) between PBV determinations was calculated voxel-wise in lung subvolumes.Main results.Of this cohort of 19 patients, 17 had a DECT acquisition at baseline. PBV maps determined from the commercial software and the TMD method were very strongly correlated [rs(PBVsyngoR=1,PBVTMDL=1) = 0.94 ± 0.01 andrs(PBVsyngoR=10,PBVTMDL=9) = 0.94 ± 0.02].PBVPETwas strongly correlated withPBVTMDL[rs(PBVPET,PBVTMDL=28) = 0.67 ± 0.11]. Perfusion patterns differed along the posterior-anterior direction [rs(PBVPET,PBVTMDL=28) = 0.77 ± 0.13/0.57 ± 0.16 in the anterior/posterior region].Significance. A strong correlation between DECT and PET determination of PBV was observed. Streak and smoothing effects in DECT and gravitational artefacts and misregistration in PET reduced the correlation posteriorly.

Investigating the role of delayed contrast magnetic resonance imaging (MRI) to differentiate radiation necrosis from tumour recurrence in brain metastases after stereotactic radiosurgery.

INTRODUCTION: The incidence of radionecrosis (RN) after stereotactic radiosurgery (SRS) to brain metastases is increasing. An overlap in the conventional MRI appearances of RN and tumour recurrence (TR) is diagnostically challenging. Delayed contrast MRI compares contrast enhancement over two time periods to create treatment response assessment maps (TRAMs). We aim to assess the utility of TRAMs in brain metastases patients. METHODS: Delayed contrast MRI scans were performed on ten brain metastases patients, previously treated with SRS, who developed equivocal lesion(s) on routine MRI follow-up. T1-weighted images were obtained five minutes and 60-75 min after contrast injection, followed by Brain Lab software analysis to create TRAMs. TRAMs patterns were then compared with the patient's clinical status, subsequent imaging, and histology results. RESULTS: We identified three regions on TRAMs: central, peripheral, and surrounding. Each region could be described either as contrast accumulation (red colour and representing non-tumour tissue) or contrast clearance (blue colour and representing tumour tissue). Our analysis demonstrated similarities in the TRAMs pattern between TR and RN, though to varying degrees. CONCLUSION: In conclusion, the TRAMs appearances of RN and TR overlap. Our findings suggest that the previously-described correlation between contrast clearance and TR is at least partially attributable to more solid initial enhancement, rather than convincingly a difference in the underlying tissue properties, and the additional diagnostic value of TRAMs may be limited. Thus, further research on TRAMs is necessary prior to incorporating it into routine clinical management after SRS for brain metastases.

The impact of inter-observer variation in delineation on robustness of radiomics features in non-small cell lung cancer.

Artificial intelligence and radiomics have the potential to revolutionise cancer prognostication and personalised treatment. Manual outlining of the tumour volume for extraction of radiomics features (RF) is a subjective process. This study investigates robustness of RF to inter-observer variation (IOV) in contouring in lung cancer. We utilised two public imaging datasets: 'NSCLC-Radiomics' and 'NSCLC-Radiomics-Interobserver1' ('Interobserver'). For 'NSCLC-Radiomics', we created an additional set of manual contours for 92 patients, and for 'Interobserver', there were five manual and five semi-automated contours available for 20 patients. Dice coefficients (DC) were calculated for contours. 1113 RF were extracted including shape, first order and texture features. Intraclass correlation coefficient (ICC) was computed to assess robustness of RF to IOV. Cox regression analysis for overall survival (OS) was performed with a previously published radiomics signature. The median DC ranged from 0.81 ('NSCLC-Radiomics') to 0.85 ('Interobserver'-semi-automated). The median ICC for the 'NSCLC-Radiomics', 'Interobserver' (manual) and 'Interobserver' (semi-automated) were 0.90, 0.88 and 0.93 respectively. The ICC varied by feature type and was lower for first order and gray level co-occurrence matrix (GLCM) features. Shape features had a lower median ICC in the 'NSCLC-Radiomics' dataset compared to the 'Interobserver' dataset. Survival analysis showed similar separation of curves for three of four RF apart from 'original_shape_Compactness2', a feature with low ICC (0.61). The majority of RF are robust to IOV, with first order, GLCM and shape features being the least robust. Semi-automated contouring improves feature stability. Decreased robustness of a feature is significant as it may impact upon the features' prognostic capability.

Please Place Your Seat in the Full Upright Position: A Technical Framework for Landing Upright Radiation Therapy in the 21st Century.

Delivering radiotherapy to patients in an upright position can allow for increased patient comfort, reduction in normal tissue irradiation, or reduction of machine size and complexity. This paper gives an overview of the requirements for the delivery of contemporary arc and modulated radiation therapy to upright patients. We explore i) patient positioning and immobilization, ii) simulation imaging, iii) treatment planning and iv) online setup and image guidance. Treatment chairs have been designed to reproducibly position seated patients for treatment and can be augmented by several existing immobilisation systems or promising emerging technologies such as soft robotics. There are few solutions for acquiring CT images for upright patients, however, cone beam computed tomography (CBCT) scans of upright patients can be produced using the imaging capabilities of standard Linacs combined with an additional patient rotation device. While these images will require corrections to make them appropriate for treatment planning, several methods indicate the viability of this approach. Treatment planning is largely unchanged apart from translating gantry rotation to patient rotation, allowing for a fixed beam with a patient rotating relative to it. Rotation can be provided by a turntable during treatment delivery. Imaging the patient with the same machinery as used in treatment could be advantageous for online plan adaption. While the current focus is using clinical linacs in existing facilities, developments in this area could also extend to lower-cost and mobile linacs and heavy ion therapy.

Cascaded deep learning-based auto-segmentation for head and neck cancer patients: Organs at risk on T2-weighted magnetic resonance imaging.

PURPOSE: To investigate multiple deep learning methods for automated segmentation (auto-segmentation) of the parotid glands, submandibular glands, and level II and level III lymph nodes on magnetic resonance imaging (MRI). Outlining radiosensitive organs on images used to assist radiation therapy (radiotherapy) of patients with head and neck cancer (HNC) is a time-consuming task, in which variability between observers may directly impact on patient treatment outcomes. Auto-segmentation on computed tomography imaging has been shown to result in significant time reductions and more consistent outlines of the organs at risk. METHODS: Three convolutional neural network (CNN)-based auto-segmentation architectures were developed using manual segmentations and T2-weighted MRI images provided from the American Association of Physicists in Medicine (AAPM) radiotherapy MRI auto-contouring (RT-MAC) challenge dataset (n = 31). Auto-segmentation performance was evaluated with segmentation similarity and surface distance metrics on the RT-MAC dataset with institutional manual segmentations (n = 10). The generalizability of the auto-segmentation methods was assessed on an institutional MRI dataset (n = 10). RESULTS: Auto-segmentation performance on the RT-MAC images with institutional segmentations was higher than previously reported MRI methods for the parotid glands (Dice: 0.860 ± 0.067, mean surface distance [MSD]: 1.33 ± 0.40 mm) and the first report of MRI performance for submandibular glands (Dice: 0.830 ± 0.032, MSD: 1.16 ± 0.47 mm). We demonstrate that high-resolution auto-segmentations with improved geometric accuracy can be generated for the parotid and submandibular glands by cascading a localizer CNN and a cropped high-resolution CNN. Improved MSDs were observed between automatic and manual segmentations of the submandibular glands when a low-resolution auto-segmentation was used as prior knowledge in the second-stage CNN. Reduced auto-segmentation performance was observed on our institutional MRI dataset when trained on external RT-MAC images; only the parotid gland auto-segmentations were considered clinically feasible for manual correction (Dice: 0.775 ± 0.105, MSD: 1.20 ± 0.60 mm). CONCLUSIONS: This work demonstrates that CNNs are a suitable method to auto-segment the parotid and submandibular glands on MRI images of patients with HNC, and that cascaded CNNs can generate high-resolution segmentations with improved geometric accuracy. Deep learning methods may be suitable for auto-segmentation of the parotid glands on T2-weighted MRI images from different scanners, but further work is required to improve the performance and generalizability of these methods for auto-segmentation of the submandibular glands and lymph nodes.

CT slice alignment to whole-body reference geometry by convolutional neural network.

Volumetric medical imaging lacks a standardised coordinate geometry which links image frame-of-reference to specific anatomical regions. This results in an inability to locate anatomy in medical images without visual assessment and precludes a variety of image analysis tasks which could benefit from a standardised, machine-readable coordinate system. In this work, a proposed geometric system that scales based on patient size is described and applied to a variety of cases in computed tomography imaging. Subsequently, a convolutional neural network is trained to associate axial slice CT image appearance with the standardised coordinate value along the patient superior-inferior axis. The trained neural network showed an accuracy of ± 12 mm in the ability to predict per-slice reference location and was relatively stable across all annotated regions ranging from brain to thighs. A version of the trained model along with scripts to perform network training in other applications are made available. Finally, a selection of potential use applications are illustrated including organ localisation, image registration initialisation, and scan length determination for auditing diagnostic reference levels.

Radiomics feature stability of open-source software evaluated on apparent diffusion coefficient maps in head and neck cancer.

Radiomics is a promising technique for discovering image based biomarkers of therapy response in cancer. Reproducibility of radiomics features is a known issue that is addressed by the image biomarker standardisation initiative (IBSI), but it remains challenging to interpret previously published radiomics signatures. This study investigates the reproducibility of radiomics features calculated with two widely used radiomics software packages (IBEX, MaZda) in comparison to an IBSI compliant software package (PyRadiomics). Intensity histogram, shape and textural features were extracted from 334 diffusion weighted magnetic resonance images of 59 head and neck cancer (HNC) patients from the PREDICT-HN observational radiotherapy study. Based on name and linear correlation, PyRadiomics shares 83 features with IBEX and 49 features with MaZda, a sub-set of well correlated features are considered reproducible (IBEX: 15 features, MaZda: 18 features). We explore the impact of including non-reproducible radiomics features in a HNC radiotherapy response model. It is possible to classify equivalent patient groups using radiomic features from either software, but only when restricting the model to reliable features using a correlation threshold method. This is relevant for clinical biomarker validation trials as it provides a framework to assess the reproducibility of reported radiomic signatures from existing trials.

A systematic review and meta-analysis of the prognostic value of radiomics based models in non-small cell lung cancer treated with curative radiotherapy.

BACKGROUND AND PURPOSE: Radiomics allows extraction of quantifiable features from imaging. This study performs a systematic review and meta-analysis of the performance of radiomics based prognostic models in non-small cell lung cancer (NSCLC). MATERIALS AND METHODS: A literature review was performed following PRISMA guidelines. Medline, EMBASE and Cochrane databases were searched for articles investigating radiomics features predictive of overall survival (OS) in NSCLC treated with curative intent radiotherapy. A random-effects meta-analysis of Harrell's Concordance Index (C-index) was performed on the performance of radiomics models. RESULTS: Of the 2746 articles retrieved, 40 studies of 55 datasets and 6223 patients were eligible for inclusion in the systematic review. There was significant heterogeneity in the methodology for feature selection and model development. Twelve datasets reported the C-index of radiomics based models in predicting OS and were included in the meta-analysis. The C-index random effects estimate was 0.57 (95% CI 0.53-0.62). There was significant heterogeneity (I2 = 70.3%). CONCLUSIONS: Based on this review, radiomics based models for lung cancer have to date demonstrated modest prognostic capabilities. Future research should consider using standardised radiomics features, robust feature selection and model development, and deep learning techniques, absolving the need for pre-defined features, to improve imaging-based models.

The effect of stereotactic body radiotherapy (SBRT) using flattening filter-free beams on cardiac implantable electronic devices (CIEDs) in clinical situations.

PURPOSE: This study focused on determining risks from stereotactic radiotherapy using flattening filter-free (FFF) beams for patients with cardiac implantable electronic device (CIEDs). Two strategies were employed: a) a retrospective analysis of patients with CIEDs who underwent stereotactic radiosurgery (SRS)/SBRT at the Peter MacCallum Cancer Centre between 2014 and 2018 and b) an experimental study on the impact of FFF beams on CIEDs. METHODS: A retrospective review was performed. Subsequently, a phantom study was performed using 30 fully functional explanted CIEDs from two different manufacturers. Irradiation was carried out in a slab phantom with 6-MV and 10-MV FFF beams. First, a repetition-rate test (RRT) with a range of beam pulse frequencies was conducted. Then, multifraction SBRT (48 Gy/4 Fx) and single-fraction SBRT (28 Gy/1 Fx) treatment plans were used for lung tumors delivered to the phantom. RESULTS: Between 2014 and 2018, 13 cases were treated with an FFF beam (6 MV, 1400 MU/min or 10 MV, 2400 MU/min), and 15 cases were treated with a flattening filter (FF) beam (6 MV, 600 MU/min). All the devices were positioned outside the treatment field at a distance of more than 5 cm, except for one case, and no failures were reported due to SBRT/SRS. In the phantom rep-rate tests, inappropriate sensing occurred, starting at a rep-rate of 1200 MU/min. Cardiac implantable electronic device anomalies during and after delivering VMAT-SBRT with a 10-MV FFF beam were observed. CONCLUSIONS: The study showed that caution should be paid to managing CIED patients when they undergo SBRT using FFF beams, as it is recommended by AAPM TG-203. Correspondingly, it was found that for FFF beams although there is small risk from dose-rate effects, delivering high dose of radiation with beam energy greater than 6 MV and high-dose rate to CIEDs positioned in close vicinity of the PTV may present issues.

Upright Radiation Therapy-A Historical Reflection and Opportunities for Future Applications.

Since the early days of megavoltage Radiation Therapy (RT), the potential of delivering treatment to a sub group of patients in an upright position has been recognized. Compared to lying horizontally, treating patients in an upright position offers potential benefits in terms of patient comfort especially for patients experiencing dyspnoea and saliva accumulation when lying down. Dosimetric benefits can also be gained from changes in the volume and location of lungs and heart in an upright position, which are potentially advantageous for clinical situations including Hodgkin's disease, lung and breast malignancies. Since the 1950's, upright stabilization mechanisms have ranged from standalone chair based apparatus to couch-top attachments with increasingly customizable solutions. The introduction of Computed-Tomography (CT) based three-dimensional (3D) dosimetry in the 1980's-90's necessitated image acquisition in a horizontal position (supine or prone), significantly reducing options for alternative patient positioning and upright techniques. Despite this, upright techniques have still been utilized where clinically indicated for palliative and novel approaches often involving non-standard treatment scenarios. More recently, a small number of centers have reported on specialized equipment capable of acquiring planning data with the patient in a vertical position. The possibility of acquiring planning quality Cone Beam CT (CBCT) on linear accelerators has recently reinvigorated the potential to deliver highly accurate and targeted treatments to patients in an upright position. This paper reflects on the historical applications of upright RT and explores new possibilities for this technology in modern RT departments.

Spin Lock Adiabatic Correction (SLAC) for B1-insensitive pulse design at 7T.

A new framework for B1 insensitive adiabatic pulse design is proposed, denoted Spin Lock Adiabatic Correction (SLAC), which counteracts deviations from ideal behaviour through inclusion of an additional correction component during pulse design. SLAC pulses are theoretically derived, then applied to the design of enhanced BIR-4 and hyperbolic secant pulses to demonstrate practical utility of the new pulses. At 7T, SLAC pulses are shown to improve the flip angle homogeneity compared to a standard adiabatic pulse with validation in both simulations and phantom experiments, under SAR equivalent experimental conditions. The SLAC framework can be applied to any arbitrary adiabatic pulse to deliver excitation with increased B1 insensitivity.