Gdp Hmm Aapmchallenge
Gdp Hmm Aapmchallenge is the code base for Generalizable Dose Prediction for Heterogenous Multi-Cohort and Multi-Site Radiotherapy Planning (GDP-HMM) challenge at AAPM 2025
Last verified:
What is Gdp Hmm Aapmchallenge?
The Gdp Hmm Aapmchallenge is an AI competition organized by AAPM (American Association of Physicists in Medicine) focused on generalizable dose prediction for radiotherapy. Participants develop fast and robust AI models to predict 3D dose distributions in radiotherapy planning across diverse tumor sites, delivery techniques, and prescription doses.
Key features include access to a dataset of over 3,500 radiotherapy plans covering head-and-neck and lung sites with IMRT and VMAT planning modes—more than 10x larger than the current largest cleaned public dataset. The challenge runs in three phases: Phase I (training dataset available), Phase II (validation datasets), and Phase III (final test datasets). Participants submit solutions via Docker containers to the Codabench platform for evaluation.
This challenge is designed for AI researchers, medical physicists, radiotherapy planners, and machine learning engineers working on automated radiotherapy planning. Top performers receive recognition including complimentary registration to present at the AAPM Annual Meeting & Exhibition, co-authorship on a summary journal paper, and acknowledgment during the Awards & Honors Ceremony.
Gdp Hmm Aapmchallenge pricing
Pricing model: Freemium
Free to participate. The challenge is sponsored by AAPM and Siemens Healthineers. Training dataset, validation dataset, baseline code, and tutorials are provided at no cost. Top two teams receive complimentary registration to the 2025 AAPM Annual Meeting & Exhibition in Washington, DC.
Gdp Hmm Aapmchallenge pros
- Large dataset with over 3,500 RT plans for robust AI training
- Covers both head-and-neck and lung tumor sites
- Includes both IMRT and VMAT planning modes
- More than 10x larger than previous public datasets
- Three-phase challenge structure for gradual model development
- Public training split with inputs and labels
- Docker-based submission ensures reproducibility
- Public leaderboard for performance tracking
- Top teams get complimentary AAPM meeting registration
- Co-authorship opportunity on summary journal paper
- Baseline code and tutorials available on GitHub
- Numpy format data works directly with baseline code
- Input data includes CT images, PTVs, OARs, beam geometries
- Validated by experts from Siemens Healthineers, Varian, UCSD, UPenn, MSKCC
- Post-challenge evaluation support with leaderboard maintained
Gdp Hmm Aapmchallenge cons
- Non-commercial use only (research purposes only)
- Maximum 3 submissions allowed in final testing phase
- Cannot delete submitted results once submitted
- Plans in dataset are not clinical-approved
- Validation split only has input shared publicly
- Test split remains fully hidden even post-challenge
- Must register challenge before downloading data
- Requires Docker installation for submissions
- Some structures not optimized with objectives
Frequently asked questions about Gdp Hmm Aapmchallenge
What is the main objective of the GDP-HMM Challenge?
The objective is to develop fast and robust AI models for generalizable 3D dose prediction in radiotherapy, and validate the effectiveness with reference high-quality plans and deliverable planning across diverse tumor sites, delivery techniques, and prescription doses.
How many radiotherapy plans are in the dataset?
The dataset includes over 3,500 RT plans (some sources say over 3,700), covering head-and-neck and lung sites with IMRT and VMAT planning modes. This is more than 10x larger than the current largest cleaned public dataset.
What are the challenge phases and dates?
Phase I starts in January 2025 with training dataset and script availability. Phase II starts February 15, 2025 when validation datasets are released. Phase III starts April 25, 2025 with final test datasets. The testing phase deadline is May 13, 2025, with results announced May 20, 2025.
How do I submit my solution to the challenge?
Participants must submit their solutions via Docker container through the Codabench platform. You need to register the challenge first under 'My Submissions', then upload your Docker container image as a tar file. Maximum 3 algorithms can be maintained per team.
Can I use this dataset for commercial purposes?
No. The dataset is for research only and commercial use is not allowed. By downloading the dataset before the challenge ends, you agree to non-commercial use only and must register for the GDP-HMM challenge.
What input data is included in the training split?
The training split includes CT images, PTVs (Planning Target Volumes), OARs (Organs at Risk), helper structures, beam geometries, prescribed dose, and corresponding dose labels for supervised learning.
What awards do top teams receive?
Top five teams get awarded and co-authored in the summary journal paper. Top two teams (one representative per team) receive complimentary registration to attend and present at the 2025 AAPM Annual Meeting & Exhibition in Washington, DC, July 27-30, 2025.
Is the test dataset publicly available after the challenge?
No. The test split remains fully hidden even after the challenge. Post-challenge, researchers can contact the lead organizer ([email protected]) for collaboration to test on the hidden split. Reference plans of the validation set will be released after the challenge.
What GitHub resources are available for participants?
A GitHub repository contains baseline code, pre-trained models, and tutorials to help participants get started. The repo provides data in Numpy format that works directly with the baseline. Raw DICOM format data is available separately at Radiotherapy_HaN_Lung_AIRTP.
What are the limitations on submissions?
Teams cannot delete submitted results. After submitting three times in the final testing phase, no further submissions are allowed. Each team can maintain up to 3 live algorithms on the platform, and updates require replacing existing algorithm slots.