Welcome to NODE21

As of April 2022, NODE21 reopens for public participation (please see "Phases" section at the end of this page)

NODE21 is a challenge with two tracks. We invite teams to develop algorithms to detect nodules and algorithms to generate nodules in chest radiographs. The goal is to develop, collectively, a high-performance open-source solution for this important clinical task. We will use the algorithms to generate nodules to create additional training data for the detection algorithms. We hope to show that generative models can be used to improve detection models for a relevant medical AI application.

Why?

Among both men and women, lung cancer causes the greatest number of cancer deaths worldwide. Symptoms of lung cancer typically occur at an advanced stage of the disease, when treatment has a reduced chance of success. Early detection is therefore a key factor in reducing mortality rates from lung cancer. Pulmonary nodules, detected through imaging, are the initial manifestation of lung cancer, visible well before clinical symptoms or signs emerge. They can be visible on a chest radiograph (CXR), and chest radiography is by far the most common radiological exam in the world. Thus, CXR plays a critical role in the accurate identification of nodules in the drive towards early detection of lung cancer. Pulmonary nodules are frequently encountered as incidental findings in patients undergoing routine examination or CXR imaging for issues unrelated to lung cancer.

It's not easy

The detection of lung nodules on CXR can be difficult, depending on their size, density, and location. Since the CXR is a projection image, the nodule is projected to the same pixels as other anatomical structures including the heart, hilum, or diaphragm, which can make them extremely subtle or even impossible to detect. This is illustrated in the figure below.

Left: a coronal CT slice from a patient with a 2 cm squamous cell lung cancer (red box). On CT, this tumor is impossible to miss. Middle: this is the average of CT slices covering 10 cm of tissue centered on the slice on the left. The nodule is still clearly visible. Right: this is the average of all CT slices and similar to what a chest radiograph of this patient would look like. Surprisingly, even though this is a large nodule, not obscured by major organs like the heart, it is still only faintly visible due to all the other structures superimposed on it. This makes nodule detection from chest radiographs challenging.

There are commercial software products on the market to assist the radiologist in identifying the locations of nodules on CXR. However, the benchmarks used for the performance of these products are not publicly available. In the research community, particularly in the era of deep learning, there has been relatively little focus on accurately detecting and localizing nodules on CXR. We believe this is partly due to the high cost of annotating nodule locations on large sets of data.

How?

Anybody can participate in NODE21, in the generation or the detection track, or in both. Unlike many other challenges, NODE21 is not set up as a competition, but rather as a collaboration. The goal is to collectively develop an open-source solution for nodule detection on chest x-rays. You can participate by putting your code in a public github repository, from which a docker container can be built to create an algorithm on the grand-challenge platform. Full details and a template for how to structure the code are provided. Your repository should contain a permissive open-source license.

We provide baseline algorithms for both generation and detection tracks. The code structure can be used by the participants as a template for their own repositories. We encourage participants to adapt these templates to run their own method or to model their docker image on them for submission to NODE21.

Detection track algorithms should read a frontal CXR, and return a list of possible bounding boxes for nodules, with a likelihood score for each bounding box.

Generation track algorithms should also input a frontal chest radiograph and additionally a json file with locations where nodules should be generated. They should produce an image with a generated nodule at the requested location. More details regarding the algorithms can be found on the Details page.

We will, together with the participants, design various protocols where systems are trained using a mix of images with real and generated nodules. Solutions are available as open-source code and as Algorithms that can be used at scale by any user of grand-challenge.org.

Phases

In early 2022 we completed a 2 phase challenge with prize winners in each track, and creators of interesting algorithms invited to co-author a publication on the topic. This publication is currently in preparation. As of April 2022 the NODE21 challenge re-opens to the general public and continues to record the current state of the art in this field.

🏆Prize

Three top-performing solutions from each track will be rewarded. These algorithms must improve upon the performance of the baseline algorithms provided by the organizers and must include a short article describing the methodology. The authors of the top-performing algorithms will receive AWS credits as follows.

~~🥇1st place in each track: AWS credits to the value of 5000 USD.~~

~~🥈2nd place in each track: AWS credits to the value of 3000 USD.~~

~~🥉3rd place in each track: AWS credits to the value of 2000 USD.~~

🏆Authorship

We plan to invite at least three solutions from each track based on their final performance, their methodology and the write-up provided by the authors for inclusion in a peer-reviewed article about the challenge. For this article, we will include additional experiments. These methods will be selected to ensure diversity of methodology as well as excellent performance. The authors of the selected algorithms (maximum of three authors per algorithm) will be invited to be a co-author of the NODE21 overview article which will be submitted to a high-impact journal.