Click here to


Are you sure ?

Yes, do it No, cancel

Spatial, Anatomically-Localized Mapping of Dose-Toxicity Associations Following Prostate Radiotherapy

M Marcello1 , A Kennedy2 , J Dowling3 , A Haworth4 , L Holloway5 , S Gulliford6 , D Dearnaley7 , J Denham8 , M Ebert9*, (1) University of Western Australia, Perth, Western Australia, (2) Sir Charles Gairdner Hospital, Perth, 6009, (3) CSIRO, Brisbane, Queensland, (4) University of Sydney, Sydney, New South Wales, (5) Liverpool and Macarthur cancer therapy centres and ingham Institute, Sydney, NSW, (6) Institute of Cancer Research & Royal Marsden, Sutton, ,(7) Institute for Cancer Research, Sutton, Surrey, (8) University of Newcastle, Newcastle, New South Wales, (9) The University of Western Australia, Nedlands,


(Monday, 7/30/2018) 4:30 PM - 6:00 PM

Room: Davidson Ballroom A

Purpose: To examine anatomically-localized associations between radiation delivery and toxicity incidence for prostate radiotherapy patients.

Methods: Treatment planning and toxicity outcomes (median > 60 months) for 750 patients were available, treated with a variety of conformal techniques ensuring diversity in dose distributions and relatively high toxicity rates. CT images were deformably registered to all members of a reference set and unsupervised clustering used to select a reference patient from the set on the basis of similarity. All dose distributions were deformed onto that reference to allow voxel-level associations between dose and outcome. Machine learning classification and Cox proportional hazards modelling methods were used to examine the relationships between planned dose and multiple gastrointestinal (GI) and genitourinary (GU) outcomes – “toxicity maps�. The robustness of predicted toxicity maps was evaluated by repeating the process using different reference patients.

Results: Classification models to predict specific toxicities performed poorly with the best-performing model based on a support vector machine with linear kernel (area under the receiver operating characteristic (AUC) of 0.67). The same models were however able to classify dose distributions according to number of treatment beams with AUC > 0.99. Voxel-wise proportional hazards modelling indicated multiple anatomical regions with significant association with toxicity outcomes, with the most striking results occurring for GU symptoms. Dysuria correlates highly (hazard ratios 1.5 – 2.2) with dose at all points along the urethra from prostatic apex to penile bulb. Haematuria correlates highly with doses around the bladder wall and particularly posteriorly (hazard ratios 3 – 5). The hazard ratios decreased and became spatially more diffuse when using a less-similar reference patient.

Conclusion: Classification models were unable to discriminate toxicity outcomes on the basis of dose distribution. Voxel-level association however indicated multiple areas of dose-toxicity association, with many of these agreeing with likely toxicity pathology.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by the Australian National Health and Medical Research Council (APP1077788).


Modeling, Prostate Therapy, Radiosensitivity


TH- response assessment : Machine learning

Contact Email