Data:

How to access dataset

1.  Complete the Google Form to access the Github where you can download the hidden Rad dataset. (Google Form)

2.  Then we will send the access right.

3.  For the sample study, the first release data sample is open in the public github directory:

• for task 1, go to https://github.com/hidden-rad/Task1. (MIMIC license is required.)

• for task 2, go to https://github.com/hidden-rad/Task2. 

Interpretable dataset

Core of the data is based on a real doctor's diagnosis process.

All output data corresponding to the report was generated through GPT-4o from answers obtained through questions according to the diagnostic procedure, and validated by radiologist review for task 1 case. You can see below what QA's information includes.

• Task 1 dataset: link

• Task 2 dataset: link

Interpretable dataset annotation procedure

This dataset includes anatomical information and causative outcomes of lesions based on textbook from the medical licensing examination.

This dataset was annotated in the form of QA with the doctor's interpretation of the diagnostic procedure. The annotation procedure can be seen in the picture above. The questions consisted of five questions and mimicked the process of a real doctor making a diagnosis on a radiography.

• Step1 - Q1: Create first impressions of images for assumed disease.

• Step2 – Q2: Anatomical location tracking of abnormal finding.

• Step3 – Q3: Thoracic spine levels tracking of abnormal finding.

• Step4 – Q4: Create diagnosis of images.

• Step5 – Q5: Confirmation for checklist on the cause of diagnosis.

• Step6 – Q4: Correct diagnosis after cause confirmation. (Checklist consists of 28 questions per diagnosis)

• Step7 – Review: Cross review. (compare the MIMIC-CXR report with the annotation)

Comparison with related datasets

• No dataset containing causes and consequences for lesions. 

• This dataset was augmented based on MIMIC-CXR.