Survival prognostication in esophageal cancer using deep learning based segmentation features
Leroy Volmer,
The Netherlands
PD-0171
Abstract
Survival prognostication in esophageal cancer using deep learning based segmentation features
Authors: Leroy Volmer1, Vasco Prudente2,1, Zhen Zhang3, Andre Dekker1,3, Maaike Berbee1, Leonard Wee1,4
1Radiation Oncology (MAASTRO), Maastricht University Medical Centre, Maastricht, The Netherlands; 2Maastricht University, CARIM School for Cardiovascular Diseases, Maastricht, The Netherlands; 3Maastricht University, GROW School of Oncology, Maastricht, The Netherlands; 4Maastricht University, GROW School of Oncology , Maastricht, The Netherlands
Show Affiliations
Hide Affiliations
Purpose or Objective
Esophageal cancer (EC) has a poor prognosis (estimated 1-year overall survival of approximately 50%). Individualized treatment selection is complicated by multiple treatment options and a lack of reliable prognostic models . Radiomics is a promising opportunity for analysis, but requires a delineation of the Gross Tumour Volume (GTV) by a physician. This study examines (a) the accuracy of automated segmentation of the GTV on planning CT series using deep learning and (b) whether deep radiomics features within a segmentation network might have prognostic potential for overall survival (OS).
Material and Methods
Clinical features and planning CTs were extracted for 406 subjects single Dutch institution treated by either definitive (DCRT) or neoadjuvant (NACRT) chemoradiotherapy for locally advanced EC. These were randomly split into training (n=325) and validation (n=81). An independent test set comprised of 52 subjects previously treated in a clinical trial (CROSS ). Each GTV was defined by a qualified radiotherapy oncologist and taken “as-treated” from the RT plan.
A squeeze and excitation segmentation network was trained to automatically segment GTV using an Adam optimizer and a compound (Focal and Dice) loss function. Deep features were extracted from intermediate convolutional layers using global average pooling, and used as predictors in a deep learning time-to-event network for OS. The Dice similarity coefficient (DSC) was used to quantify geometric segmentation agreement, and Harrell’s concordance index (HCI) was used to evaluate discriminative performance. Confidence intervals were estimated via bootstrap resampling.
Results
In the test set, the mean DSC was 0.72 (range 0.26-0.86). The HCI for OS at 2 years were: 0.65 (95% confidence interval 0.64-0.65), 0.66 (0.66-0.67) and 0.68 (0.68-0.69), for deep features only, clinical features only (age, sex, tumour location, treatment intent, T-stage, N-stage and tumour volume) and a combination of (clinical + deep) features, respectively.
Conclusion
Deep learning automatic segmentation was able to make a reasonable estimate of the GTV on hitherto unseen planning CT series. The deep features from segmentation taken to a deep learning survival model appear to contain prognostic information regarding OS. The combination of clinical and deep features leads to a small increase in discriminative performance.