Feb 15, 2017
Shrinkage predictions boost lung tumour dose
Radiotherapy is the standard of care for individuals with inoperable, locally advanced non-small cell lung cancer (NSCLC). However, local control of the disease is only achieved in 30 to 40% of patients. Attempts to improve disease control by escalating tumour dose are hampered by often large tumour volumes that sit within normal lung and close to sensitive structures such as the spinal cord and heart.
Tackling this, researchers at the Memorial Sloan Kettering Cancer Center have developed an atlas – a population-based model – that predicts tumour shrinkage over the course of treatment. In a retrospective study using so-called predictive treatment planning (PTP), they concurrently escalated dose to a predicted residual tumour volume while limiting dose to organs-at-risk (OARs). In creating a single treatment plan, the approach avoids time-consuming replans in the middle of a patient's treatment course (Phys. Med. Biol. 62 702).
The atlas uses a database of clinical tumour shrinkage data provided by CT and weekly on-treatment cone-beam CT (CBCT) scans of a group of patients who had already been treated. Tumour shrinkage in an incoming patient is predicted using their planning CT scan and a probability map that combines the shrinkage patterns observed in the database, transformed onto the patient's geometry.
Shrinkage patterns of the database patients are characterized using principal component analysis. This approach models tumour geometry more accurately than a prototype model previously reported by the researchers (Plan incorporates tumour shrinkage). There, pre- and post-treatment tumours were approximated as ellipsoids and shrinkage was modelled as an affine transformation between the two, a technique that does not preserve details such as concave surfaces.
Other improvements included an increase in database size from five to 12 individuals. The new atlas also accounts for the different shrinkage patterns observed in parenchymal tumours and those attached to the chest wall or mediastinum. While parenchymal tumours tend to shrink towards their centre, tumours attached to the mediastinum or chest wall shrink asymmetrically towards the anchoring structures.
The researchers assessed the new atlas by planning treatments for the 12 patients in the database. Each patient's tumour shrinkage was predicted using a database of the 11 remaining cases. The plans concurrently escalated dose to the predicted residual tumour planning target volume, while delivering the regular prescribed dose to the clinical PTV. At the same time, OAR doses were restricted to the limits specified in the clinical protocol used at the centre.
The dose received by 95% of the PTV of the actual residual tumour, D95 PTVresd, was 6.5 ± 1.8 Gy higher than that in the clinical, non-escalated plan. The mean dose to the same volume was 10.4 ± 2.9 Gy higher. The researchers were also able to steer dose hot spots into the residual volumes. Dose escalation was not as high, however, as in plans generated using the actual residual tumour, due to differences between the predicted and actual volumes. Using the predicted volume, D95 PTVresd was 2.9 ± 1.6 Gy lower and the mean dose was 0.9 ± 0.5 Gy lower.
Based on their latest results, the researchers are optimistic of the atlas' ultimate clinical potential. "We will be able to maximize the local-regional control of NSCLC by selectively escalating dose to the resistant sub-volumes of the tumour while limiting the toxicity to OAR within the standard clinical guidelines," said first author Perry Zhang.
In ongoing work, the researchers are working to predict tumour shrinkage more accurately by adding more patients to the database and incorporating a supporting vector machine, a machine learning technique. "The biggest challenge is to find patients with proper longitudinal imaging studies and secure enough patients for the creation of the atlas," said Zhang. The authors have since recruited over 30 patients in collaboration with researchers at the Netherlands Cancer Institute (NKI) in Amsterdam. A clinical trial in which the model will be used to prospectively plan treatments is planned for within the next two years.
About the author
Jude Dineley is a freelance science writer and former medical physicist based in Bavaria, Germany.