About DRAW
An open-source autosegmentation system for radiotherapy
DRAW was developed to lower the barriers to adopting automatic segmentation in radiotherapy, particularly in low- and middle-income countries where commercial systems are often prohibitively expensive and difficult to customise.
The Challenge in Radiotherapy Contouring
Modern precision radiotherapy depends on accurate segmentation of target volumes and organs at risk. This process is time-consuming and subject to significant inter- and intra-observer variability.
Automatic segmentation saves valuable time and additionally reduces variability in contouring. However, the key challenges in adopting commercial automatic segmentation systems include high cost, vendor lock-in, a lack of representation of the local population in the training dataset, and limited capability to customise automatic segmentation models.
In low- and middle-income countries (LMIC) with limited resources and technical know-how, deploying these models is not trivial. Most centres lack the technical know-how and resources to deploy inference pipelines in their local settings, even after purchasing adequate computing hardware. Furthermore, contouring practices vary across centres, which means hospitals want a customisable system and the ability to determine how they want a specific volume delineated.
Automatic segmentation saves valuable time and additionally reduces variability in contouring. However, the key challenges in adopting commercial automatic segmentation systems include high cost, vendor lock-in, a lack of representation of the local population in the training dataset, and limited capability to customise automatic segmentation models.
In low- and middle-income countries (LMIC) with limited resources and technical know-how, deploying these models is not trivial. Most centres lack the technical know-how and resources to deploy inference pipelines in their local settings, even after purchasing adequate computing hardware. Furthermore, contouring practices vary across centres, which means hospitals want a customisable system and the ability to determine how they want a specific volume delineated.
Our Approach
We set up the DRAW project in 2024 to develop a deep learning-based radiotherapy automatic segmentation workflow. The system was first deployed at a tertiary cancer centre in an urban area of Eastern India (Tata Medical Center, Kolkata). Initial results from a prostate cancer autosegmentation model were promising, and we soon expanded to include gynaecological cancers, head and neck cancer, brain tumours, and breast cancers.
The initial deployment relied on a local GPU workstation. After discussions with colleagues at other oncology centres, we realised most centres lacked the resources to deploy inference pipelines locally. In 2025, we redesigned the architecture to address these requirements.
To be cost-efficient and lower the barrier to entry, we opted for a client-server architecture in which multiple thin clients upload DICOM data to a centralised cloud server for automatic segmentation. To maintain patient privacy, DICOM data is de-identified before export. The inference layer was decoupled from the server, enabling on-demand use of expensive, energy-consuming GPU resources.
The initial deployment relied on a local GPU workstation. After discussions with colleagues at other oncology centres, we realised most centres lacked the resources to deploy inference pipelines locally. In 2025, we redesigned the architecture to address these requirements.
To be cost-efficient and lower the barrier to entry, we opted for a client-server architecture in which multiple thin clients upload DICOM data to a centralised cloud server for automatic segmentation. To maintain patient privacy, DICOM data is de-identified before export. The inference layer was decoupled from the server, enabling on-demand use of expensive, energy-consuming GPU resources.
Architecture Overview
The DRAW system comprises three components: a client, an API server, and ephemeral GPU instances which are created on demand.
1. DRAW Client — Deployed on any modern desktop operating system using Docker. Minimum requirements: 64-bit processor, 2GB RAM, 10GB disk space. It supports asynchronous task processing using Celery and serves as a DICOM server using Pynetdicom. It supports two mechanisms for receiving DICOM data: a DICOM SCP (C-Store Service Class Provider) for automatic receiving, or polling a network directory. Once installed, it automatically processes incoming DICOM images, matches segmentation templates via rulesets, de-identifies per HIPAA Safe Harbor, and uploads to the API server.
2. DRAW API Server — Deployed on a VPS with 2 vCPUs, 8 GB RAM, 100 GB disk. Provides model provisioning services and task orchestration. Authenticates each transfer using client-specific JWT access tokens stored encrypted at rest. Verifies images, sends them to ephemeral GPU instances, and retrieves the RTStructureSet after segmentation. Also deletes de-identified data after successful transfer.
3. Inference Pipeline — Deployed on a VPS with 1 GPU (typically 8-12 GB VRAM) using a preconfigured OS template. Runs nnU-Net v2 inference on ephemeral containers. The client and server are decoupled from the inference layer, so more advanced segmentation frameworks can be used in the future.
1. DRAW Client — Deployed on any modern desktop operating system using Docker. Minimum requirements: 64-bit processor, 2GB RAM, 10GB disk space. It supports asynchronous task processing using Celery and serves as a DICOM server using Pynetdicom. It supports two mechanisms for receiving DICOM data: a DICOM SCP (C-Store Service Class Provider) for automatic receiving, or polling a network directory. Once installed, it automatically processes incoming DICOM images, matches segmentation templates via rulesets, de-identifies per HIPAA Safe Harbor, and uploads to the API server.
2. DRAW API Server — Deployed on a VPS with 2 vCPUs, 8 GB RAM, 100 GB disk. Provides model provisioning services and task orchestration. Authenticates each transfer using client-specific JWT access tokens stored encrypted at rest. Verifies images, sends them to ephemeral GPU instances, and retrieves the RTStructureSet after segmentation. Also deletes de-identified data after successful transfer.
3. Inference Pipeline — Deployed on a VPS with 1 GPU (typically 8-12 GB VRAM) using a preconfigured OS template. Runs nnU-Net v2 inference on ephemeral containers. The client and server are decoupled from the inference layer, so more advanced segmentation frameworks can be used in the future.
Deployment & Impact
The DRAW system was made operational on April 9th, 2025 at Centre A (Tata Medical Center, Kolkata) and on April 23rd, 2025 at Centre B (Mohananda Cancer Centre, Durgapur — a charitable cancer centre in a Tier II city in Eastern India).
Between April 2025 and April 2026, 4215 of 4303 uploaded CT scan series (98%) were successfully segmented. The API server uptime exceeded 99.5%. The median segmentation processing time was 20.2 minutes (IQR: 16.9-28.9 minutes).
The system covers anatomical organs, clinical target volumes (both nodal and primary), and planning structures for treatment planning. Cancer sites covered include breast, head and neck, lung, brain, prostate, cervix, and rectum.
The system is vendor-agnostic: Centre A uses Varian and Accuray treatment planning, while Centre B uses Elekta. The output can be used irrespective of the treatment planning system vendor.
Between April 2025 and April 2026, 4215 of 4303 uploaded CT scan series (98%) were successfully segmented. The API server uptime exceeded 99.5%. The median segmentation processing time was 20.2 minutes (IQR: 16.9-28.9 minutes).
The system covers anatomical organs, clinical target volumes (both nodal and primary), and planning structures for treatment planning. Cancer sites covered include breast, head and neck, lung, brain, prostate, cervix, and rectum.
The system is vendor-agnostic: Centre A uses Varian and Accuray treatment planning, while Centre B uses Elekta. The output can be used irrespective of the treatment planning system vendor.
Deployment Sites MapImage to be added
Team & Partners
DRAW is a collaboration between:
Department of Radiation Oncology, Tata Medical Center, Kolkata, India
Department of Computer Science, Indian Institute of Technology, Kharagpur, India
The system has been deployed at two centres: Tata Medical Center, Kolkata (Centre A) and Mohananda Cancer Centre, Durgapur (Centre B, a charitable cancer centre in a Tier II city in Eastern India).
The project brings together clinical expertise in radiation oncology with deep-learning engineering to build practical, deployable autosegmentation tools for real-world cancer care, particularly in low- and middle-income settings.
Department of Radiation Oncology, Tata Medical Center, Kolkata, India
Department of Computer Science, Indian Institute of Technology, Kharagpur, India
The system has been deployed at two centres: Tata Medical Center, Kolkata (Centre A) and Mohananda Cancer Centre, Durgapur (Centre B, a charitable cancer centre in a Tier II city in Eastern India).
The project brings together clinical expertise in radiation oncology with deep-learning engineering to build practical, deployable autosegmentation tools for real-world cancer care, particularly in low- and middle-income settings.
Want to learn more?
Read the detailed system architecture and data flow, or register your interest.