Here we made the C-SLAKE publicly. The motivation for this dataset is to ensure the robustness of our model . We extend the publicly available Med-VQA dataset Slake to a consistent Med-VQA dataset C-SLAKE for further consistency assessment.
The dadaset can be used for general Med-VQA task, and also for testing model consistency.
This is the folder structure of the dataset:
📦C-SLAKE
┣📂imgs
┃┗Contains images and corresponding question and answer pairs.
┣📂train
┃┗Contains train question answer pairs.
┣📂validate
┃┗Contains validate question answer pairs.
┣📂test
┃┗Contains test question answer pairs.
You can download the image from the above link and put it in the C-SLAKE folder.
The questions in our dataset are selected from the Slake dataset. We conducted role-tag annotation humanly to establish connections between interconnected question-answer pairs corresponding to the identified medical images. Diagnostic question-answer pairs are categorized into three distinct roles. The distribution of role-type questions is presented in the following table.
Statistics of question in different role types over training, validation, and testing sets:
Role-type | train | Val | Test | Total |
---|---|---|---|---|
Main | 1222 | 237 | 246 | 1705 |
Context | 814 | 158 | 170 | 1142 |
Ind | 2294 | 514 | 508 | 3316 |
Overall | 4330 | 909 | 924 | 6163 |
The C-SLAKE dataset encompasses a diverse array of medical materials,including Computed Tomography (CT) scans, Magnetic Resonance Imaging (MRI)scans, and X-Ray images. The distribution of it is presented in the following table:
The distribution of diverse array of medical materials
Modality-type | nums |
---|---|
CT | 282 |
MRI | 181 |
X-Ray | 179 |
Overall | 642 |
This dataset was extended by Slake. We thank the original authors for their work and open source dataset.