This challenge is hosted on Kaggle InClass: https://www.kaggle.com/c/pet-radiomics-challenges
Cancers arising from the head and neck have become increasingly more studied in the past few years, especially cancers of the oropharynx which are now epidemic domestically, with over 20,000 annual cases projected in the U.S. Growing evidence has established differential clinical, pathological, molecular, and epidemiological attributes between human papilloma virus-associated (HPV+) and HPV-negative (HPV-) oropharyngeal cancer (OPC) disease entities.
Recent data suggest that "radiomics", or extraction of image texture analysis to generate mineable quantitative data from medical images, can reflect phenotypes for various cancers. Several groups have shown that developed radiomics signatures, in head and neck cancers among other tumor sites, can be correlated with survival outcomes. Radiomics features from metabolic imaging modalities, like 18F-fluorodeoxyglucose Positron Emission Tomography (18F-FDG PET) have been postulated as surrogates for underlying tumor biology, and hence prognosis.
This competition is organized as a Medical Image Computing and Computer Assisted Intervention (MICCAI) Computational Precision Medicine (CPM) grand challenge. Contestants are tasked to predict, using primary tumor 18F-FDG PET-derived radiomics features +/- matched clinical data, whether a tumor arising from the oropharynx will be controlled by definitive radiation treatment (RT). The head and neck radiation oncology team from University of Texas MD Anderson Cancer Center (MDACC) have curated and harmonized a multi-institutional dataset of 248 oropharynx cancer (OPC) patients, using our in-house 'LAMBDA-RAD' data management platform. Scans came from six different institutes from: the US (MDACC), Canada [four different clinical institutions in Québec: Hôpital Général Juif de Montréal (HGJ), Centre Hospitalier Universitaire de Sherbrooke (CHUS), Centre Hopitalier de l'Université de Montréal (CHUM) and Hôpital Maisonneuve-Rosemont de Montréal (HMR)], and Europe (MAASTRO Clinic, The Netherlands).
This dataset encompasses anonymized 18F-FDG-PET DICOM-RT files, and expert-segmented contours of primary tumors propagated from computed tomography (CT) to 18F-FDG-PET scans [DICOM-RTSTRUCT]. This is further supplemented with relevant clinical data, known etiological/biological correlates and tumor control outcomes as ground truth. Our major target is to assess the ability of participant-developed radiomics workflows to predict binary (phenotypic/genotypic) (local tumor recurrence) using a defined “Training” cohort as a prior complete dataset that includes all input and outcome data, to build up an algorithm.
The human factors that mimic real life clinical practice should be taken into consideration in a way that participants should foresee encountering: tumor segmentation and image co-registration, as well as inter-scanner variability.
Participants will be ranked according the the 'Matthews correlation coefficient' calculated in the private subset of the test set.
Specifically, we will post one-half of the 18F-FDG-PET files from the dataset, in DICOM-RT format, on the Kaggle InClass server system, as a “training set”. Challenge participants will also be able to download a “test set", including the half of the remaining dataset (public subset of test set.
Relevant etiological, clinical and treatment-related details are supplied for both sets, with the expected outcome values, namely the "local tumor control", being available only in the training set.
Contestants are invited to download the DICOM-RT files, along with clinical meta-data tables, with subsequent mechanistic analysis,that includes the performance of individual risk assessments. The region of interest for robust texture analysis should be the primary gross tumor volume, which is denoted as GTVp. The ultimate goal will be the development of an algorithm that yields the probability of local tumor control in oropharynx cancer patients who received definitive radiation treatment. This model MUST be based on GTVp-derived radiomics features +/- associated clinical data.
The performance of the contestants’ models will be first assessed on the public subset of the test set and results will be posted to a ‘Public leaderboard’. This will provide real-time feedback to contestants on the performance of their models relative to that of other contestants’ models.
Toward the end of the challenge, each contestant/team will be allowed to select his/her/their own two ‘optimal’ final submissions of choice. Contestants will then be judged according to the performance of their chosen model(s) on the hold-out subset of test set, according to the private leaderboard. More details can be found on the challenge page on Kaggle InClass https://www.kaggle.com/c/pet-radiomics-challenges#Evaluation
In order to reinforce the "FAIR Guiding Principles for scientific data management and stewardship" (doi:10.1038/sdata.2016.18), we request the following from the top-3 winners:
1) The source code used to select image features, train and test prediction models be sent to the challenge organizers on the day following the end of the test phase (winners will be contacted by the organizers); and
2) That source code to be made available online (e.g. on GitHub) prior to the MICCAI 2018 meeting.
Participants failing to follow these rules will be rejected from the top-3 winner list, and the next best ranking participants will be contacted to complete the list.More details can be found at: https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=37224869
3. Submit an extended abstract (up to 4-pages) before the end of the test phase describing their algorithm. Please upload the abstract report on dropbox folder
Start: June 15, 2018, 11:59 p.m.
Start: July 31, 2018, midnight
You must be logged in to participate in competitions.Sign In