Feasibility of case-based beam generation for robotic radiosurgery

AI-based optimization for US-guided radiation therapy of the prostate

OBJECTIVES

Fast volumetric ultrasound presents an interesting modality for continuous and real-time intra-fractional target tracking in radiation therapy of lesions in the abdomen. However, the placement of the ultrasound probe close to the target structures leads to blocking some beam directions.

METHODS

To handle the combinatorial complexity of searching for the ultrasound-robot pose and the subset of optimal treatment beams, we combine CNN-based candidate beam selection with simulated annealing for setup optimization of the ultrasound robot, and linear optimization for treatment plan optimization into an AI-based approach. For 50 prostate cases previously treated with the CyberKnife, we study setup and treatment plan optimization when including robotic ultrasound guidance.

RESULTS

The CNN-based search substantially outperforms previous randomized heuristics, increasing coverage from 93.66 to 97.20% on average. Moreover, in some cases the total MU was also reduced, particularly for smaller target volumes. Results after AI-based optimization are similar for treatment plans with and without beam blocking due to ultrasound guidance.

CONCLUSIONS

AI-based optimization allows for fast and effective search for configurations for robotic ultrasound-guided radiation therapy. The negative impact of the ultrasound robot on the plan quality can successfully be mitigated resulting only in minor differences.

Feasibility and analysis of CNN-based candidate beam generation for robotic radiosurgery

PURPOSE

Robotic radiosurgery offers the flexibility of a robotic arm to enable high conformity to the target and a steep dose gradient. However, treatment planning becomes a computationally challenging task as the search space for potential beam directions for dose delivery is arbitrarily large. We propose an approach based on deep learning to improve the search for treatment beams.

METHODS

In clinical practice, a set of candidate beams generated by a randomized heuristic forms the basis for treatment planning. We use a convolutional neural network to identify promising candidate beams. Using radiological features of the patient, we predict the influence of a candidate beam on the delivered dose individually and let this prediction guide the selection of candidate beams. Features are represented as projections of the organ structures which are relevant during planning. Solutions to the inverse planning problem are generated for random and CNN-predicted candidate beams.

RESULTS

The coverage increases from 95.35% to 97.67% for 6000 heuristically and CNN-generated candidate beams, respectively. Conversely, a similar coverage can be achieved for treatment plans with half the number of candidate beams. This results in a patient-dependent reduced averaged computation time of 20.28%-45.69%. The number of active treatment beams can be reduced by 11.35% on average, which reduces treatment time. Constraining the maximum number of candidate beams per beam node can further improve the average coverage by 0.75 percentage points for 6000 candidate beams.

CONCLUSIONS

We show that deep learning based on radiological features can substantially improve treatment plan quality, reduce computation runtime, and treatment time compared to the heuristic approach used in clinics.

Methods for a similarity measure for clinical attributes based on survival data analysis

Background

Case-based reasoning is a proven method that relies on learned cases from the past for decision support of a new case. The accuracy of such a system depends on the applied similarity measure, which quantifies the similarity between two cases. This work proposes a collection of methods for similarity measures especially for comparison of clinical cases based on survival data, as they are available for example from clinical trials.

Methods

Our approach is intended to be used in scenarios, where it is of interest to use longitudinal data, such as survival data, for a case-based reasoning approach. This might be especially important, where uncertainty about the ideal therapy decision exists. The collection of methods consists of definitions of the local similarity of nominal as well as numeric attributes, a calculation of attribute weights, a feature selection method and finally a global similarity measure. All of them use survival time (consisting of survival status and overall survival) as a reference of similarity. As a baseline, we calculate a survival function for each value of any given clinical attribute.

Results

We define the similarity between values of the same attribute by putting the estimated survival functions in relation to each other. Finally, we quantify the similarity by determining the area between corresponding curves of survival functions. The proposed global similarity measure is designed especially for cases from randomized clinical trials or other collections of clinical data with survival information. Overall survival can be considered as an eligible and alternative solution for similarity calculations. It is especially useful, when similarity measures that depend on the classic solution-describing attribute “applied therapy” are not applicable. This is often the case for data from clinical trials containing randomized arms.

Conclusions

In silico evaluation scenarios showed that the mean accuracy of biomarker detection in k = 10 most similar cases is higher (0.909–0.998) than for competing similarity measures, such as Heterogeneous Euclidian-Overlap Metric (0.657–0.831) and Discretized Value Difference Metric (0.535–0.671). The weight calculation method showed a more than six times (6.59–6.95) higher weight for biomarker attributes over non-biomarker attributes. These results suggest that the similarity measure described here is suitable for applications based on survival data.

Treatment Planning Considerations for Robotic Guided Cardiac Radiosurgery for Atrial Fibrillation

Purpose

Robotic guided stereotactic radiosurgery has recently been investigated for the treatment of atrial fibrillation (AF). Before moving into human treatments, multiple implications for treatment planning given a potential target tracking approach have to be considered.   

Materials & Methods

Theoretical AF radiosurgery treatment plans for twenty-four patients were generated for baseline comparison. Eighteen patients were investigated under ideal tracking conditions, twelve patients under regional dose rate (RDR = applied dose over a certain time window) optimized conditions (beam delivery sequence sorting according to regional beam targeting), four patients under ultrasound tracking conditions (beam block of the ultrasound probe) and four patients with temporary single fiducial tracking conditions (differential surrogate-to-target respiratory and cardiac motion).  

Results

With currently known guidelines on dose limitations of critical structures, treatment planning for AF radiosurgery with 25 Gy under ideal tracking conditions with a 3 mm safety margin may only be feasible in less than 40% of the patients due to the unfavorable esophagus and bronchial tree location relative to the left atrial antrum (target area). Beam delivery sequence sorting showed a large increase in RDR coverage (% of voxels having a larger dose rate for a given time window) of 10.8-92.4% (median, 38.0%) for a 40-50 min time window, which may be significant for non-malignant targets. For ultrasound tracking, blocking beams through the ultrasound probe was found to have no visible impact on plan quality given previous optimal ultrasound window estimation for the planning CT. For fiducial tracking in the right atrial septum, the differential motion may reduce target coverage by up to -24.9% which could be reduced to a median of -0.8% (maximum, -12.0%) by using 4D dose optimization. The cardiac motion was also found to have an impact on the dose distribution, at the anterior left atrial wall; however, the results need to be verified.

Conclusion

Robotic AF radiosurgery with 25 Gy may be feasible in a subgroup of patients under ideal tracking conditions. Ultrasound tracking was found to have the lowest impact on treatment planning and given its real-time imaging capability should be considered for AF robotic radiosurgery. Nevertheless, advanced treatment planning using RDR or 4D respiratory and cardiac dose optimization may be still advised despite using ideal tracking methods.

Inverse treatment planning for spinal robotic radiosurgery: an international multi-institutional benchmark trial

Stereotactic radiosurgery (SRS) is the accurate, conformal delivery of high‐dose radiation to well‐defined targets while minimizing normal structure doses via steep dose gradients. While inverse treatment planning (ITP) with computerized optimization algorithms are routine, many aspects of the planning process remain user‐dependent. We performed an international, multi‐institutional benchmark trial to study planning variability and to analyze preferable ITP practice for spinal robotic radiosurgery. 10 SRS treatment plans were generated for a complex‐shaped spinal metastasis with 21 Gy in 3 fractions and tight constraints for spinal cord (V14Gy<2 cc, V18Gy<0.1 cc) and target (coverage >95%). The resulting plans were rated on a scale from 1 to 4 (excellent‐poor) in five categories (constraint compliance, optimization goals, low‐dose regions, ITP complexity, and clinical acceptability) by a blinded review panel. Additionally, the plans were mathematically rated based on plan indices (critical structure and target doses, conformity, monitor units, normal tissue complication probability, and treatment time) and compared to the human rankings. The treatment plans and the reviewers' rankings varied substantially among the participating centers. The average mean overall rank was 2.4 (1.2‐4.0) and 8/10 plans were rated excellent in at least one category by at least one reviewer. The mathematical rankings agreed with the mean overall human rankings in 9/10 cases pointing toward the possibility for sole mathematical plan quality comparison. The final rankings revealed that a plan with a well‐balanced trade‐off among all planning objectives was preferred for treatment by most participants, reviewers, and the mathematical ranking system. Furthermore, this plan was generated with simple planning techniques. Our multi‐institutional planning study found wide variability in ITP approaches for spinal robotic radiosurgery. The participants', reviewers', and mathematical match on preferable treatment plans and ITP techniques indicate that agreement on treatment planning and plan quality can be reached for spinal robotic radiosurgery.

PACS number(s): 87.55.de

Knowledge-light adaptation approaches in case-based reasoning for radiotherapy treatment planning

OBJECTIVE

Radiotherapy treatment planning aims at delivering a sufficient radiation dose to cancerous tumour cells while sparing healthy organs in the tumour-surrounding area. It is a time-consuming trial-and-error process that requires the expertise of a group of medical experts including oncologists and medical physicists and can take from 2 to 3h to a few days. Our objective is to improve the performance of our previously built case-based reasoning (CBR) system for brain tumour radiotherapy treatment planning. In this system, a treatment plan for a new patient is retrieved from a case base containing patient cases treated in the past and their treatment plans. However, this system does not perform any adaptation, which is needed to account for any difference between the new and retrieved cases. Generally, the adaptation phase is considered to be intrinsically knowledge-intensive and domain-dependent. Therefore, an adaptation often requires a large amount of domain-specific knowledge, which can be difficult to acquire and often is not readily available. In this study, we investigate approaches to adaptation that do not require much domain knowledge, referred to as knowledge-light adaptation.

METHODOLOGY

We developed two adaptation approaches: adaptation based on machine-learning tools and adaptation-guided retrieval. They were used to adapt the beam number and beam angles suggested in the retrieved case. Two machine-learning tools, neural networks and naive Bayes classifier, were used in the adaptation to learn how the difference in attribute values between the retrieved and new cases affects the output of these two cases. The adaptation-guided retrieval takes into consideration not only the similarity between the new and retrieved cases, but also how to adapt the retrieved case.

RESULTS

The research was carried out in collaboration with medical physicists at the Nottingham University Hospitals NHS Trust, City Hospital Campus, UK. All experiments were performed using real-world brain cancer patient cases treated with three-dimensional (3D)-conformal radiotherapy. Neural networks-based adaptation improved the success rate of the CBR system with no adaptation by 12%. However, naive Bayes classifier did not improve the current retrieval results as it did not consider the interplay among attributes. The adaptation-guided retrieval of the case for beam number improved the success rate of the CBR system by 29%. However, it did not demonstrate good performance for the beam angle adaptation. Its success rate was 29% versus 39% when no adaptation was performed.

CONCLUSIONS

The obtained empirical results demonstrate that the proposed adaptation methods improve the performance of the existing CBR system in recommending the number of beams to use. However, we also conclude that to be effective, the proposed adaptation of beam angles requires a large number of relevant cases in the case base.

Similar-case-based optimization of beam arrangements in stereotactic body radiotherapy for assisting treatment planners

Objective. To develop a similar-case-based optimization method for beam arrangements in lung stereotactic body radiotherapy (SBRT) to assist treatment planners. Methods. First, cases that are similar to an objective case were automatically selected based on geometrical features related to a planning target volume (PTV) location, PTV shape, lung size, and spinal cord position. Second, initial beam arrangements were determined by registration of similar cases with the objective case using a linear registration technique. Finally, beam directions of the objective case were locally optimized based on the cost function, which takes into account the radiation absorption in normal tissues and organs at risk. The proposed method was evaluated with 10 test cases and a treatment planning database including 81 cases, by using 11 planning evaluation indices such as tumor control probability and normal tissue complication probability (NTCP). Results. The procedure for the local optimization of beam arrangements improved the quality of treatment plans with significant differences (P < 0.05) in the homogeneity index and conformity index for the PTV, V10, V20, mean dose, and NTCP for the lung. Conclusion. The proposed method could be usable as a computer-aided treatment planning tool for the determination of beam arrangements in SBRT.