Automatic segmentation is gaining increasing acceptance in Radiation Oncology to reduce burden and improve consistency. However commercial systems are expensive and inflexible. Open source systems are flexible but require a significant technological know-how for implementation. The team from Tata Medical Center Kolkata and IIT Kharagpur have developed a deep-learning based autosegmentation system which can be deployed with minimal capital expenditure and uses a distributed client-server architecture to allow fast parallel processing of cases. At the time of this version, the DRAW system uses nnU-Net for the deep-learning-based modelling.

Aims
To evaluate the real world use of the  DRAW system in a real world setting

Objectives
To determine if the DRAW system can be successfully installed and configured at multiple hospitals (Implementation objective)
To determine if the system performance is considered acceptable for the users at the hospitals where the DRAW system was installed (Validation objective)

Endpoints
To determine the percentage of centers where the DRAW system was successfully installed and configured. The DRAW client system should be installed and work with the DICOM data storage system available at the participating center to be considered as a successful install. 
Qualitative: At least 50% of the cases were segmented with no or minor modifications/errors as determined qualitatively by the end users.
Quantitative: At least 70% of the segmented structures have an average volumetric dice similarity score (VDS) not less than 0.10 points below the reference VDS. The reference value will be defined based on the VDS reported in the literature (or if not available in the literature then the model validation statistics obtained from the DRAW segmentation pipeline). 

Design
Multicenter, single-arm, prospective cohort study

Methods
The DRAW system will be installed in the participating hospitals and configured to integrate it with the planning workflow. Autosegmentation will be performed by the DRAW system and this will be reviewed by the oncologists at the participating center. Qualitative evaluation of the contours will be performed by the oncologists. De-identified DICOM data will be used to determine the spatial similarity between autosegmented and manually segmented structures for quantitative evaluation.

Statistical Analysis
Descriptive summary statistics will be presented for each endpoint along with appropriate visualization. For proportions we will also report the corresponding binomial 95% confidence intervals. For quantitative data, the median, mean and 95% confidence intervals of the mean shall be reported.  

Sample size
The minimum sample size for the study is 1000 patients based on the quantitative validation objective. 

Study Duration
2 years