Natural language processing and machine learning to assist radiation oncology incident learning.

PURPOSE: To develop a Natural Language Processing (NLP) and Machine Learning (ML) pipeline that can be integrated into an Incident Learning System (ILS) to assist radiation oncology incident learning by semi-automating incident classification. Our goal was to develop ML models that can generate label recommendations, arranged according to their likelihoods, for three data elements in Canadian NSIR-RT taxonomy. METHODS: Over 6000 incident reports were gathered from the Canadian national ILS as well as our local ILS database. Incident descriptions from these reports were processed using various NLP techniques. The processed data with the expert-generated labels were used to train and evaluate over 500 multi-output ML algorithms. The top three models were identified and tuned for each of three different taxonomy data elements, namely: (1) process step where the incident occurred, (2) problem type of the incident and (3) the contributing factors of the incident. The best-performing model after tuning was identified for each data element and tested on unseen data. RESULTS: The MultiOutputRegressor extended Linear SVR models performed best on the three data elements. On testing, our models ranked the most appropriate label 1.48 ± 0.03, 1.73 ± 0.05 and 2.66 ± 0.08 for process-step, problem-type and contributing factors respectively. CONCLUSIONS: We developed NLP-ML models that can perform incident classification. These models will be integrated into our ILS to generate a drop-down menu. This semi-automated feature has the potential to improve the usability, accuracy and efficiency of our radiation oncology ILS.

Development and implementation of a radiation therapy incident learning system compatible with local workflow and a national taxonomy.

PURPOSE: Collaborative incident learning initiatives in radiation therapy promise to improve and standardize the quality of care provided by participating institutions. However, the software interfaces provided with such initiatives must accommodate all participants and thus are not optimized for the workflows of individual radiation therapy centers. This article describes the development and implementation of a radiation therapy incident learning system that is optimized for a clinical workflow and uses the taxonomy of the Canadian National System for Incident Reporting - Radiation Treatment (NSIR-RT). METHODS: The described incident learning system is a novel version of an open-source software called the Safety and Incident Learning System (SaILS). A needs assessment was conducted prior to development to ensure SaILS (a) was intuitive and efficient (b) met changing staff needs and (c) accommodated revisions to NSIR-RT. The core functionality of SaILS includes incident reporting, investigations, tracking, and data visualization. Postlaunch modifications of SaILS were informed by discussion and a survey of radiation therapy staff. RESULTS: There were 240 incidents detected and reported using SaILS in 2016 and the number of incidents per month tended to increase throughout the year. An increase in incident reporting occurred after switching to fully online incident reporting from an initial hybrid paper-electronic system. Incident templating functionality and a connection with our center's oncology information system were incorporated into the investigation interface to minimize repetitive data entry. A taskable actions feature was also incorporated to document outcomes of incident reports and has since been utilized for 36% of reported incidents. CONCLUSIONS: Use of SaILS and the NSIR-RT taxonomy has improved the structure of, and staff engagement with, incident learning in our center. Software and workflow modifications informed by staff feedback improved the utility of SaILS and yielded an efficient and transparent solution to categorize incidents with the NSIR-RT taxonomy.

Polymer segregation under confinement: free energy calculations and segregation dynamics simulations.

Monte Carlo simulations are used to study the behavior of two polymers under confinement in a cylindrical tube. Each polymer is modeled as a chain of hard spheres. We measure the free energy of the system, F, as a function of the distance between the centers of mass of the polymers, λ, and examine the effects on the free energy functions of varying the channel diameter D and length L, as well as the polymer length N and bending rigidity κ. For infinitely long cylinders, F is a maximum at λ = 0, and decreases with λ until the polymers are no longer in contact. For flexible chains (κ = 0), the polymers overlap along the cylinder for low λ, while above some critical value of λ they are longitudinally compressed and non-overlapping while still in contact. We find that the free energy barrier height, ΔF ≡ F(0) - F(∞), scales as ΔF/k(B)T ∼ ND(-1.93 ± 0.01), for N ⩽ 200 and D ⩽ 9σ, where σ is the monomer diameter. In addition, the overlap free energy appears to scale as F/k(B)T = Nf(λ/N; D) for sufficiently large N, where f is a function parameterized by the cylinder diameter D. For channels of finite length, the free energy barrier height increases with increasing confinement aspect ratio L/D at fixed volume fraction ϕ, and it decreases with increasing ϕ at fixed L/D. Increasing the polymer bending rigidity κ monotonically reduces the overlap free energy. For strongly confined systems, where the chain persistence length P satisfies D ≪ P, F varies linearly with λ with a slope that scales as F'(λ) ∼ -k(B)TD(-ß)P(-α), where ß ≈ 2 and α ≈ 0.37 for N = 200 chains. These exponent values deviate slightly from those predicted using a simple model, possibly due to insufficiently satisfying the conditions defining the Odijk regime. Finally, we use Monte Carlo dynamics simulations to examine polymer segregation dynamics for fully flexible chains and observe segregation rates that decrease with decreasing entropic force magnitude, f ≡ |dF/dλ|. For both infinite-length and finite-length channels, the polymers are not conformationally relaxed at later times during segregation.

Analysis of multicatheter interstitial brachytherapy: Accelerated partial breast irradiation in a retrospective cohort of early-stage breast cancer patients.

PURPOSE: To determine cardiac dose received by patients treated with high dose rate interstitial brachytherapy. Patients with early-stage, node negative breast cancer can be treated using multi-catheter interstitial brachytherapy accelerated partial breast irradiation (MIB-APBI), with the benefit of reduced treatment volumes and favorable toxicity. METHODS AND MATERIALS: We conducted a retrospective review of left-sided breast cancer patients treated using MIB-APBI at our institution since 2014. The mean heart dose (MHD) was calculated using the Oncentra 3.2 planning system. The minimum distance between the planning target volume (PTVeval) and heart contour was measured manually. FINDINGS: 81 patients were included. The upper outer quadrant was the most common site. The MHD was 97.8 cGy (EQD2a/b=2) (range 22-229 cGy). MHD significantly correlated with the closest distance between PTVeval and heart contour (correlation coefficient -0.823, p <0.001); size of PTVeval (cc) and quadrant location did not. CONCLUSIONS: Appropriately selected women with early-stage, low-risk, left-sided breast cancer who received MIB-APBI had acceptable MHD. There was a strong correlation between the distance of PTVeval and MHD. Quadrant breast tumor is in cannot be used as a surrogate for MHD in brachytherapy. Our findings contribute to the growing evidence of the utility and safety of MIB-APBI.

A Monte Carlo study on the impact of indirect action on neutron relative biological effectiveness.

Recent Monte Carlo studies have linked the energy-dependent risk of neutron-induced stochastic effects to the relative biological effectiveness (RBE) of neutrons in inflicting difficult-to-repair clusters of lesions in nuclear deoxyribonucleic acid (DNA). However, an investigation on the damaging effects of indirect radiation action is missing from such studies. In this work, we extended our group's existing simulation pipeline by incorporating and validating a model for indirect action. Our updated simulation pipeline was used to study the impact of indirect action and estimate neutron RBE for inflicting clustered lesions in DNA. In our results, although indirect action significantly increased the average yield of DNA damage clusters, our neutron RBE values are lower in magnitude than previous estimates due to model limitations and the greater relative impact of indirect action in lower-linear energy transfer (LET) radiation than in higher-LET radiation.

A study of indirect action's impact on simulated neutron-induced DNA damage.

Objective.The risk of radiobiological stochastic effects associated with neutrons is strongly energy dependent. Recent Monte Carlo studies simulating neutron-irradiated nuclear DNA have demonstrated that this energy dependence is correlated with the relative biological effectiveness (RBE) of neutrons to inflict DNA damage clusters that contain difficult-to-repair double-strand breaks. However, these previous investigations were either limited to modeling direct radiation action or considered the effects of both direct and indirect action together without distinguishing between the two. In this study, we aimed to quantify the influence of indirect action in neutron irradiation scenarios and acquire novel estimations of the energy-dependent neutron RBE for inducing DNA damage clusters due to both direct and indirect action.Approach.We explored the role of indirect action in neutron-induced DNA damage by integrating a validated indirect action model into our existing simulation pipeline. Using this pipeline, we performed track-structure simulations of monoenergetic neutron irradiations (1 eV to 10 MeV) in a nuclear DNA model and analyzed the resulting simple and clustered DNA lesions. We repeated the irradiation simulations for 250 keV x-rays that acted as our reference radiation.Main results.Including indirect action significantly increased the occurrence of DNA lesions. We found that indirect action tends to amplify the damage due to direct action by inducing DNA lesions in the vicinity of directly-induced lesions, resulting in additional and larger damage clusters. Our neutron RBE results are qualitatively similar to but lower in magnitude than the established radiation protection factors and the results of previous similar investigations, due to the greater relative impact of indirect action in photon-induced damage than in neutron-induced damage.Significance.Although our model for neutron-induced DNA damage has some important limitations, our findings suggest that the energy-dependent risk of neutron-induced stochastic effects may not be completely modeled alone by the relative potential of neutrons to inflict clustered lesions via direct and indirect action in DNA damage.

Towards the characterization of neutron carcinogenesis through direct action simulations of clustered DNA damage.

Neutron exposure poses a unique radiation protection concern because neutrons have a large, energy-dependent relative biological effectiveness (RBE) for stochastic effects. Recent computational studies on the microdosimetric properties of neutron dose deposition have implicated clustered DNA damage as a likely contributor to this marked energy dependence. So far, publications have focused solely on neutron RBE for inducing clusters of DNA damage containing two or more DNA double strand breaks (DSBs). In this study, we have conducted a novel assessment of neutron RBE for inducing all types of clustered DNA damage that contain two or more lesions, stratified by whether the clusters contain DSBs (complex DSB clusters) or not (non-DSB clusters). This assessment was conducted for eighteen initial neutron energies between 1 eV and 10 MeV as well as a reference radiation of 250 keV x-rays. We also examined the energy dependence of cluster length and cluster complexity because these factors are believed to impact the DNA repair process. To carry out our investigation, we developed a user-friendly TOPAS-nBio application that includes a custom nuclear DNA model and a novel algorithm for recording clustered DNA damage. We found that neutron RBE for inducing complex DSB clusters exhibited similar energy dependence to the canonical neutron RBE for stochastic radiobiological effects, at multiple depths in human tissue. Qualitatively similar results were obtained for non-DSB clusters, although the quantitative agreement was lower. Additionally we identified a significant neutron energy dependence in the average length and complexity of clustered lesions. These results support the idea that many types of clustered DNA damage contribute to the energy dependence of neutron RBE for stochastic radiobiological effects and imply that the size and constituent lesions of individual clusters should be taken into account when modeling DNA repair. Our results were qualitatively consistent for (i) multiple radiation doses (including a low-dose 0.1 Gy irradiation), (ii) variations in the maximal lesion separation distance used to define a cluster, and (iii) two distinct collections of physics models used to govern particle transport. Our complete TOPAS-nBio application has been released under an open-source license to enable others to independently validate our work and to expand upon it.

The impact of treatment parameter variation on secondary neutron spectra in high-energy electron beam radiotherapy.

High-energy electron treatment procedures in radiotherapy pose a potential iatrogenic cancer risk as well as a critical health risk to patients with cardiac implantable electronic devices due to the generation of secondary neutrons in the linac head, the treatment room, and the patient. It may be argued that the neutron production from photons is well characterized, but the same is not true for electrons. Therefore, to assess the risk involved in an electron treatment, one must determine the neutron flux spectrum generated by the treatment procedure. The neutron spectrum depends on the treatment parameters used and therefore it is crucial to study its dependence on these parameters. In this work, eight experiments were devised to analyze how eight electron treatment parameters impacted the neutron spectrum. The parameters we considered were the electron beam energy, location of measurement, cutout size, collimator size, applicator size, collimator angle, choice of treatment room, and the presence or absence of a solid water phantom. For each experiment, we used a Nested Neutron Spectrometer™ (NNS) to measure the neutron flux spectra for multiple settings of the treatment parameter of interest. The resulting spectra were plotted and compared. We found that the electron beam energy and the location of measurement had the most impact on the neutron flux spectra, while the other parameters had a smaller or insignificant impact. This report may serve as a reference tool for medical physicists to help estimate the risk associated with a particular high-energy electron treatment procedure.

The effect of the flattening filter on photoneutron production at 10 MV in the Varian TrueBeam linear accelerator.

PURPOSE: Neutrons are an unavoidable by-product of high-energy radiation therapy treatments that deliver unwanted nontarget dose to patients. Use of flattening-filter-free (FFF) photon beams has been shown to significantly reduce photoneutron production per monitor unit (MU) of dose delivered. The purpose of this investigation was to characterize the photoneutron production of the 10 MV and 10 MV FFF beams of the Varian TrueBeamTM linear accelerator. METHODS: Neutron fluence spectra were measured using a Nested Neutron SpectrometerTM (NNS, Detec Inc., Gatineau, Canada). The ratios of neutron fluence and ambient dose equivalent for the 10 MV FFF beam relative to the 10 MV beam, dubbed FF-ratios (FFF/FF), were used to characterize the difference between the two beams. FF-ratios were compared under the following three conditions (a) per MU, at various locations in the treatment room, (b) per MU, with the linac jaws opened and closed, and (c) per electron striking the bremsstrahlung target, as opposed to per MU, at one location with the jaws closed. RESULTS: On average, the neutron fluence for the 10 MV FFF beam was 37% lower per MU than the 10 MV beam (FF-ratio = 0.63). The FF-ratio in neutron fluence and ambient dose equivalent did not vary by much between different locations within the treatment room. However, the FF-ratio in neutron ambient dose equivalent was reduced significantly when the linac jaws were opened compared to closed, which implies that the jaws contribute more to the photoneutron spectrum of the 10 MV FFF beam than to the 10 MV beam. Finally, it was found that the 10 MV FFF beam produces more photoneutrons per electron striking the bremsstrahlung target than the 10 MV beam (FF-ratio = 2.56). CONCLUSIONS: The photoneutron fluence per MU produced by the 10 MV FFF beam is 37% lower than the 10 MV beam of a Varian TrueBeam linac. Accordingly, a reduction in neutron dose received by patients is achieved through use of the unflattened beam, provided that treatment plans for each beam require approximately the same number of MU. It was found to be instructive to compare the photoneutron yield per source electron between the two beams as it helped provide an understanding of the physics underlying photoneutron production in both beams.

Interstitial diffuse radiance spectroscopy of gold nanocages and nanorods in bulk muscle tissues.

Radiance spectroscopy was applied to the interstitial detection of localized inclusions containing Au nanocages or nanorods with various concentrations embedded in porcine muscle phantoms. The radiance was quantified using a perturbation approach, which enabled the separation of contributions from the porcine phantom and the localized inclusion, with the inclusion serving as a perturbation probe of photon distributions in the turbid medium. Positioning the inclusion at various places in the phantom allowed for tracking of photons that originated from a light source, passed through the inclusion's location, and reached a detector. The inclusions with high extinction coefficients were able to absorb nearly all photons in the range of 650-900 nm, leading to a spectrally flat radiance signal. This signal could be converted to the relative density of photons incident on the inclusion. Finally, the experimentally measured quantities were expressed via the relative perturbation and arranged into the classical Beer-Lambert law that allowed one to extract the extinction coefficients of various types of Au nanoparticles in both the transmission and back reflection geometries. It was shown that the spatial variation of perturbation could be described as 1/r dependence, where r is the distance between the inclusion and the detector. Due to a larger absorption cross section, Au nanocages produced greater perturbations than Au nanorods of equal particle concentration, indicating a better suitability of Au nanocages as contrast agents for optical measurements in turbid media. Individual measurements from different inclusions were combined into detectability maps.

Feasibility of interstitial near-infrared radiance spectroscopy platform for ex vivo canine prostate studies: optical properties extraction, hemoglobin and water concentration, and gold nanoparticles detection.

The canine prostate is a close match for the human prostate and is used in research of prostate cancers. Determining accurately optical absorption and scattering properties of the gland in a wide spectral range (preferably in a minimally invasive way), linking optical properties to concentrations of major endogenous chromophores, and detecting the presence of localized optical inhomogeneities like inclusions of gold nanoparticles for therapeutic and diagnostic purposes, are among the major challenges for researchers. The goal of the article is to demonstrate a feasibility of the multifunctional radiance spectroscopy platform in providing the required information. For ex vivo canine prostate, extraction of the effective attenuation and diffusion coefficients using relative cw radiance measurements was demonstrated in the 650- to 900-nm range. The derived absorption coefficient was decomposed to contributions from 9.0 µM HbO2, 29.6 µM Hb, and 0.47 fractional volume of H2O. Detection of a localized inclusion containing ∼1.5·1010 gold nanorods (0.8 µg Au) at 10 mm distance from the urethra was achieved with the detector in the urethra and the light source in a virtual rectum position. The platform offers the framework for a systematic study of various chromophores in the prostate that can be used as comprehensive diagnostic markers.

Optical absorption and scattering properties of bulk porcine muscle phantoms from interstitial radiance measurements in 650-900 nm range.

We demonstrated the application of relative radiance-based continuous wave (cw) measurements for recovering absorption and scattering properties (the effective attenuation coefficient, the diffusion coefficient, the absorption coefficient and the reduced scattering coefficient) of bulk porcine muscle phantoms in the 650-900 nm spectral range. Both the side-firing fiber (the detector) and the fiber with a spherical diffuser at the end (the source) were inserted interstitially at predetermined locations in the phantom. The porcine phantoms were prostate-shaped with ∼4 cm in diameter and ∼3 cm thickness and made from porcine loin or tenderloin muscles. The described method was previously validated using the diffusion approximation on simulated and experimental radiance data obtained for homogenous Intralipid-1% liquid phantom. The approach required performing measurements in two locations in the tissue with different distances to the source. Measurements were performed on 21 porcine phantoms. Spectral dependences of the effective attenuation and absorption coefficients for the loin phantom deviated from corresponding dependences for the tenderloin phantom for wavelengths <750 nm. The diffusion constant and the reduced scattering coefficient were very close for both phantom types. To quantify chromophore presence, the plot for the absorption coefficient was matched with a synthetic absorption spectrum constructed from deoxyhemoglobin, oxyhemoglobin and water. The closest match for the porcine loin spectrum was obtained with the following concentrations: 15.5 µM (±30% s.d.) Hb, 21 µM (±30% s.d.) HbO2 and 0.3 (±30% s.d.) fractional volume of water. The tenderloin absorption spectrum was best described by 30 µM Hb (±30% s.d), 19 µM (±30% s.d.) HbO2 and 0.3 (±30% s.d.) fractional volume of water. The higher concentration of Hb in tenderloin was consistent with a dark-red appearance of the tenderloin phantom. The method can be applied to a number of biological tissues and organs for interstitial optical interrogation.