Link to official GitHub repository: https://github.com/ImageCLEF/2023_ImageCLEFmed_VQA
Updates
- 28-04-2023 - Added testing dataset
- 01-03-2023 - Updated dataset with instrument segmentation masks
Identifying lesions in colonoscopy images is one of the most popular applications of artificial intelligence in medicine. Until now, the research has focused on single-image or video analysis. With this task, we aim at bringing a new aspect to the field by adding multiple modalities to the picture. The main focus of the task will be on visual question answering and visual question generation. The goal is that through the combination of text and image data the output of the analysis gets easier to use by medical experts. The task has three sub-tasks.
For the visual question answering (VQA), the participants need to combine images and text answers to answer the questions. In the visual question generation (VQG) subtask, the participants are asked to generate text questions from a given image and answer. Example questions for both VQA and VQG: How many polyps are in the image? Are there any polyps in the image? What disease is visible in the image? The third subtask is the visual location question answering (VLQA), where the participants get an image and a question and are required to answer it by providing a segmentation mask for the image. Example questions are: Where in the image is the polyp? Where in the image is the normal and the diseased part? What part of the image shows normal mucosa?
The VQA task asks participants to generate text answers given a text question and image pair. For example, we provide an image containing a colon polyp with the following question, "Where in the image is the polyp located?". Here, the answer should be a textual description of where in the image the polyp is located, like upper-left or in the center of the image.
The VQG task asks participants to generate text questions based on a given text answer and image pair. This task is an inverse of task 1, where instead of generating the answer, we are asking for the question. An example could be given the answer "The image contains a polyp" and an image containing a polyp; the question should be "Does the image contain an abnormality?".
The VLQA task asks participants to segment parts of an image given a text question and image pair. This task differs from task 1 and task 2 as it requires a visual output rather than a textual one. For example, given the question "Where is the abnormality?" and an image containing a polyp, the output should give a segmentation mask covering the polyp in the image.
The provided data is based on the Hyper Kvasir dataset datasets.simula.no/hyper-kvasir with the additional question-and-answer ground truth, which was developed with our medical partners. The data includes images spanning the entire gastrointestinal tract, from mouth to anus, and will include abnormalities, surgical instruments, and normal findings. This includes data from different procedures such as gastroscopy and colonoscopy.
For task 1 (VQA) and task 2 (VQG), we provide at 2,000 image samples. Not all questions will be relevant to the provided image, and we expect the submissions to be able to handle cases where there is no correct answer.
For task 3 (VLQA), we provide images with question and segmentation mask pairs. This part of the data will be based on the segmentation part of HyperKvasir and Kvasir Instrument.
The dataset for both task can be downlaoded below.
For assessing performance the following metrics will be used: F1 score, Accuracy, Recall, MCC, mIOU and Dice coefficient. More information to be added soon.
Please refer to the general ImageCLEF registration instructions: https://www.imageclef.org/2023.
To submit your entries for the CLEF 2023 Medical Visual Question Answering (MedVQA) Challenge, please follow these steps:
- Create a single zip file that includes a separate folder for each task you participated in.
- Name the zip file using this format:
<team_name>_clef2023_medvqa.zip. - Label the folders within the zip file as
task1,task2, etc., corresponding to the respective tasks. - Make sure to follow the detailed instructions and requirements for each task's submission, as described below.
Submissions can be emailed to the challenge organizers; Steven Hicks ([email protected]) and Michael Riegler ([email protected]).
For Task 1, your submission should include a JSON file with the following format:
- The JSON file should contain one entry per image in the test dataset.
- Each entry should provide answers to the 18 questions listed in the table below.
- Entries should be formatted as key-value pairs, with the key being the image ID and the value being a list of answers.
You can test the correctness of your submission by running the check_task_1_submission.py script on your JSON file. This script will verify the formatting and structure of your file, ensuring it meets the requirements for evaluation. You can also check the sample-submission directory for an example JSON file that meets the aforemention requirements.
| Question ID | Questions | Question ID | Questions |
|---|---|---|---|
| 0 | What type of procedure is the image taken from? | 9 | How many instruments are in the image? |
| 1 | Have all polyps been removed? | 10 | Where in the image is the abnormality? |
| 2 | Is this finding easy to detect? | 11 | Where in the image is the instrument? |
| 3 | Is there a green/black box artifact? | 12 | Are there any abnormalities in the image? |
| 4 | Is there text? | 13 | Are there any anatomical landmarks in the image? |
| 5 | What color is the abnormality? | 14 | Are there any instruments in the image? |
| 6 | What color is the anatomical landmark? | 15 | Where in the image is the anatomical landmark? |
| 7 | How many findings are present? | 16 | What is the size of the polyp? |
| 8 | How many polyps are in the image? | 17 | What type of polyp is present? |
Submissions for Task 2 are now open, and participants have the flexibility to submit a variety of materials related to their work. Some suggestions for submission content include:
- A collection of image-answer pairs, along with the generated questions associated with them.
- The source code used to create the question generation system, including any algorithms, libraries, or tools employed.
- A comprehensive report detailing the functionality and methodology of your question generation system, including any novel techniques or approaches used.
Please note that all submissions for Task 2 will be assessed through manual evaluation by a team of experts. This process will ensure a thorough and fair assessment of the quality, creativity, and effectiveness of each participant's question generation system.
Submissions for Task 3 follow a similar structure to Task 1, with a primary focus on answering questions related to segmentation. To participate in Task 3, please ensure your submission includes the following components:
-
A JSON file formatted as described in Task 1, containing answers to the segmentation-related questions listed in the table below. This file should be organized with one entry per image in the test dataset and include key-value pairs, where the key represents the image ID and the value contains a list of corresponding answers.
-
A separate directory containing the predicted segmentation masks associated with the answers provided in the JSON file. These masks should be organized and labeled in a manner that allows for easy correlation with the respective image IDs and corresponding questions.
By focusing on segmentation-related questions and providing both the JSON file and the predicted masks, participants will contribute valuable data to the evaluation of Task 3. This information will enable a comprehensive assessment of each submission's effectiveness in addressing the specific challenges associated with image segmentation in the medical domain. You can use the verify_task_3_submission.py script to check that your submission meets the requirements.
| Question ID | Questions | Question ID | Questions |
|---|---|---|---|
| 18 | Where exactly in the image is the polyp located? | 19 | Where exactly in the image is the instrument located? |
- 13 February 2023: Release of the training and validation sets
- 28 April 2023: Release of the test sets
- 28 April 2023: Registration closes
- 14 May 2023: Run submission deadline
- 17 May 2023: Release of the processed results by the task organizers
- 5 June 2023: Submission of participant papers [CEUR-WS]
- 23 June 2023: Notification of acceptance
- 7 July 2023: Camera ready copy of participant papers and extended lab overviews [CEUR-WS]
- 18-21 September CLEF Conference
- Steven A. Hicks <steven(at)simula.no>, SimulaMet, Norway
- Michael A. Riegler <michael(at)simula.no>, SimulaMet, Norway
- Vajira Thambawita <vajira(at)simula.no>, SimulaMet, Norway
- Andrea Storås <andrea(at)simula.no>, SimulaMet, Norway
- Pål Halvorsen <paalh(at)simula.no>, SimulaMet, Norway
- Thomas de Lange <thomas.de.lange(at)gu.se>, Sahlgrenska University Hospital, Sweden
- Nikolaos Papachrysos <nikolaos.papachrysos(at)vgregion.se>, Sahlgrenska University Hospital, Sweden
- Johanna Schöler <johanna.scholer(at)vgregion.se>, Sahlgrenska University Hospital, Sweden
- Debesh Jha <debesh.jha(at)northwestern.edu>, Norway & Northwestern University, USA