Introduction: A multicenter research was carried out to validate predictive strategies: to determinate patients eligible for re-planning due to predictable anatomical variations. Advantage and challenges of IGRT, deformable image registration and different set-up protocols were evaluated in a multicenter retrospective study using planning data. Materials and Methods: 76 head and neck (H&N) patients were considered with more than 2200 daily studies analyzed: 1200 MVCT from Center-A (A), 600 CBCT from Center-B (B), 240 CBCT from Center-C (C) and 240 CBCT from Center-D (D). To obtain a predictive time of warping during the 6 weeks of therapy, the study focused on volume and dose variations of parotid glands (PG). RayStation hybrid algorithm, supported by IronPython scripts and GPU computing, was used to perform a daily deformable image registration, structures automatic re-contouring and dose accumulation. A home-made machine-learning classifier tool was developed in MATLAB. Support Vector Machines (SVM) and cluster analysis were used for a weekly time-series evaluation and patients' selection. Results: At the end of the treatment (~30 fractions), the whole cohort is affected by a PG mean reduction of 23.7 ± 8.8%: 25.1 ± 9.2% (A), 23.8 ± 6.6% (B), 21.2 ± 10.3% (C) and 24.4 ± 9.8% (D). Using machine learning approach, the shrinkage of 86.3% of cases can be predicted during the first 3 weeks of therapy: 89.6% (A), 92.7% (B), 76.0% (C) and 87.0% (D). The number of patients that would benefit from a review of the initial plan reached 53.5% from the 4th week, with an inter-centers variability of 19.7%. Conclusions: A SVM and cluster decision making tool was developed and trained in order to overcome Adaptive RT logistic challenges in a busy clinical routine. The time for re-planning and the specific patients that would benefit from a new plan were quantified in this multicenter study to increase the personalization of the patients' treatment during all sessions.
A machine learning tool for re-planning and adaptive RT: A multicenter cohort investigation / Guidi, G.; Maffei, N.; Meduri, B.; Ciarmatori, A.; Mistretta, G.M.; Maggi, S.; Cardinali, M.; Morabito, V.E.; Rosica, F.; Malara, S.; Savini, A.; Orlandi, G.; D'Ugo, C.; Bunkheila, F.; Bono, M.; Lappi, S.; Blasi, C.; Giacobazzi, P.; Baldazzi, G.; Costi, T.. - In: PHYSICA MEDICA. - ISSN 1120-1797. - ELETTRONICO. - 32:Supplement 1(2016), pp. 31-32. (Intervento presentato al convegno 9th National Congress of the Associazione Italiana di Fisica Medica tenutosi a Perugia, Italy nel 25-28 February 2016) [10.1016/j.ejmp.2016.01.108].
A machine learning tool for re-planning and adaptive RT: A multicenter cohort investigation
GUIDI, GABRIELEInvestigation
;CIARMATORI, ALBERTOMembro del Collaboration Group
;BUNKHEILA, FEISAL;BLASI, CLAUDIO;BALDAZZI, GIUSEPPESupervision
;COSTI, TIZIANAMembro del Collaboration Group
2016
Abstract
Introduction: A multicenter research was carried out to validate predictive strategies: to determinate patients eligible for re-planning due to predictable anatomical variations. Advantage and challenges of IGRT, deformable image registration and different set-up protocols were evaluated in a multicenter retrospective study using planning data. Materials and Methods: 76 head and neck (H&N) patients were considered with more than 2200 daily studies analyzed: 1200 MVCT from Center-A (A), 600 CBCT from Center-B (B), 240 CBCT from Center-C (C) and 240 CBCT from Center-D (D). To obtain a predictive time of warping during the 6 weeks of therapy, the study focused on volume and dose variations of parotid glands (PG). RayStation hybrid algorithm, supported by IronPython scripts and GPU computing, was used to perform a daily deformable image registration, structures automatic re-contouring and dose accumulation. A home-made machine-learning classifier tool was developed in MATLAB. Support Vector Machines (SVM) and cluster analysis were used for a weekly time-series evaluation and patients' selection. Results: At the end of the treatment (~30 fractions), the whole cohort is affected by a PG mean reduction of 23.7 ± 8.8%: 25.1 ± 9.2% (A), 23.8 ± 6.6% (B), 21.2 ± 10.3% (C) and 24.4 ± 9.8% (D). Using machine learning approach, the shrinkage of 86.3% of cases can be predicted during the first 3 weeks of therapy: 89.6% (A), 92.7% (B), 76.0% (C) and 87.0% (D). The number of patients that would benefit from a review of the initial plan reached 53.5% from the 4th week, with an inter-centers variability of 19.7%. Conclusions: A SVM and cluster decision making tool was developed and trained in order to overcome Adaptive RT logistic challenges in a busy clinical routine. The time for re-planning and the specific patients that would benefit from a new plan were quantified in this multicenter study to increase the personalization of the patients' treatment during all sessions.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.