An image-based deep learning framework for individualising radiotherapy dose: a retrospective analysis of outcome prediction

Background: Radiotherapy continues to be delivered without consideration of individual tumour characteristics. To advance towards more precise treatments in radiotherapy, we queried the lung CT-derived feature space to identify radiation sensitivity parameters that can predict treatment failure and hence guide the individualisation of radiotherapy dose. Methods: An institutional review board-approved study (IRB 14-562) was used to identify patients treated with lung stereotactic body radiotherapy. Patients with primary (stage IA–IV) or recurrent lung cancer and patients with other cancer types with solitary metastases or oligometastases to the lung were included. Patients without digitally accessible CT image or radiotherapy structure data were excluded. The internal study cohort received treatment at the main campus of the Cleveland Clinic (Cleveland, OH, USA). The independent validation cohort received treatment at seven affiliate regional or national sites. We input pre-therapy lung CT images into Deep Profiler, a multi-task deep neural network that has radiomics incorporated into the training process, to generate an image fingerprint that predicts time-to-event treatment outcomes and approximates classical radiomic features. We validated our findings with the independent study cohort. Deep Profiler was combined with clinical variables to derive iGray, an individualised dose that estimates treatment failure probability to be below 5%. Findings: A total of 1275 patients were assessed for eligibility and 944 met our eligibility criteria; 849 were in the internal study cohort and 95 were in the independent validation cohort. Radiation treatments in patients with high Deep Profiler scores failed at a significantly higher rate than in patients with low scores; 3-year cumulative incidence of local failure in the internal study cohort was 20·3% (16·0–24·9) in patients with high Deep Profiler scores and 5·7% (95% CI 3·5–8·8) in patients with low Deep Profiler scores (hazard ratio [HR]=3·64 [95% CI 2·19–6·05], p<0·0001). Deep Profiler independently predicted local failure (HR=1·65 [1·02–2·66], p=0·042). Models that included Deep Profiler and clinical variables predicted treatment failures with a concordance index (C-index) of 0·72 (95% CI 0·67–0·77), a significant improvement compared with classical radiomics (p<0·0001) or clinical variables (p<0·0001) alone. Deep Profiler performed well in the independent validation cohort, predicting treatment failures across diverse clinical settings and CT scanner types (C-index 0·77, 95% CI 0·69–0·92). iGray had a wide dose range (21·1–277 Gy) and suggested dose reductions in 23·3% of patients. Our results also showed that iGray can be safely delivered in the majority of cases. Interpretation: Our results indicate that there are image-distinct subpopulations that have differential sensitivity to radiotherapy. The image-based deep learning framework proposed herein is the first opportunity to use medical images to individualise radiotherapy dose. Our results signify a new roadmap for deep learning-guided predictions and treatment guidance in the image-replete and highly standardised discipline of radiation oncology. Funding: Siemens Medical Solutions USA.