Wolfda95/MIRP_Your_Other_Left

Dataset of the Publication: Your other Left! Vision-Language Models Fail to Identify Relative Positions in Medical Images (MICCAI 2025)

For more Information:

• Visit the Project Page
• Or read the Paper

---

Information about the Dataset:

Each subfolder for the Research Questions (RQ1, RQ2, RQ3) and the Ablation Study (AS) contains:

• Image folders with the rotated/flipped CT slices
• JSON files with corresponding question-answer (QA) pairs for each CT slice

Example for each Research Question:

For details, check the paper summary on our Project Page.

Folder Structure of the MIRP Dataset:

📁 RQ1 (Research Question 1)

• images/ – CT slices
• qa.json – Question-answer pairs

📁 RQ2 (Research Question 2)

• image_dots/ – CT slices with dot markers
• image_letters/ – CT slices with letter markers
• image_numbers/ – CT slices with number markers
• qa_dots.json – QA pairs with anatomical names for dot markers
• qa_letters.json – QA pairs with anatomical names for letter markers
• qa_numbers.json – QA pairs with anatomical names for number markers

📁 RQ3 (Research Question 3(2))

• image_dots/ – CT slices with dot markers
• image_letters/ – CT slices with letter markers
• image_numbers/ – CT slices with number markers
• qa_dots.json – QA pairs without anatomical names for dot markers
• qa_letters.json – QA pairs without anatomical names for letter markers
• qa_numbers.json – QA pairs without anatomical names for number markers

📁 AS (Ablation Study)

• image_dots/ – White images with dot markers
• image_letters/ – White images with letter markers
• image_numbers/ – White images with number markers
• qa_dots.json – QA pairs for dot markers
• qa_letters.json – QA pairs for letter markers
• qa_numbers.json – QA pairs for number markers

Note:

• RQ3 is the dataset we used for RQ3(2).
• For RQ3(1), we use the same data as RQ1 — only the evaluation method later differs (answers are later checked for standard anatomy correctness instead of image orientation correctness).
• Need a refresher on the research questions? Check the paper summary on our Project Page.

Anatomic Structures that we use in MIRP:

We extracted the anatomical structures that the questions refer to using the Totalsegmentatior: https://github.com/wasserth/TotalSegmentator

The correct answers to the relative positioning questions are obtained by comparing the centers of mass of the two structures being evaluated.

Here is the list of all the structures (we have grouped the different vertebrae and ribs together):

Index	TotalSegmentator name
1	spleen
2	kidney_right
3	kidney_left
4	gallbladder
5	liver
6	stomach
7	pancreas
8	adrenal_gland_right
9	adrenal_gland_left
10	lung_upper_lobe_left
11	lung_lower_lobe_left
12	lung_upper_lobe_right
13	lung_middle_lobe_right
14	lung_lower_lobe_right
15	esophagus
16	trachea
17	thyroid_gland
18	small_bowel
19	duodenum
20	colon
21	urinary_bladder
22	prostate
23	kidney_cyst_left
24	kidney_cyst_right
25	sacrum
26	vertebrae
51	heart
52	aorta
53	pulmonary_vein
54	brachiocephalic_trunk
55	subclavian_artery_right
56	subclavian_artery_left
57	common_carotid_artery_right
58	common_carotid_artery_left
59	brachiocephalic_vein_left
60	brachiocephalic_vein_right
61	atrial_appendage_left
62	superior_vena_cava
63	inferior_vena_cava
64	portal_vein_and_splenic_vein
65	iliac_artery_left
66	iliac_artery_right
67	iliac_vena_left
68	iliac_vena_right
69	humerus_left
70	humerus_right
71	scapula_left
72	scapula_right
73	clavicula_left
74	clavicula_right
75	femur_left
76	femur_right
77	hip_left
78	hip_right
79	spinal_cord
80	gluteus_maximus_left
81	gluteus_maximus_right
82	gluteus_medius_left
83	gluteus_medius_right
84	gluteus_minimus_left
85	gluteus_minimus_right
86	autochthon_left
87	autochthon_right
88	iliopsoas_left
89	iliopsoas_right
90	brain
91	skull
92	rib_left
93	rib_right
116	sternum
117	costal_cartilages

Details on the Question-Answer JSON Files:

This is one exmaple element of a question answer json file.

{
  "filename": "amos_0002.nii_slice-20_classes-22_perc-12.png",
  "base_name": "amos_0002.nii",
  "slice_index": 20,
  "classes_count": 22,
  "multiple_components_same_label": true,
  "question_answer": [
    {
      "object1_name": "left kidney",
      "object2_name": "inferior vena cava",
      "object1_gray": 3,
      "object2_gray": 63,
      "object1_center_x": 346.97574777687953,
      "object1_center_y": 214.49151172190784,
      "object2_center_x": 231.29456193353474,
      "object2_center_y": 290.3836858006042,
      "question": "Is the left kidney below the inferior vena cava?",
      "answer": 0
    }
  ],
  "rotate_flip_short": "A3",
  "rotate_flip_long": "Horizontal flip and 180 degree rotation"
},

• filename: Filename of the PNG image to match the question to the image
• base_name: The original CT volume where the slice was extraced from (Either a volume from the Amos dataset or from the BTCV dataset)
• slice_index: Slice number within the CT volume.
• classes_count: Number of anatomical structures segmented by TotalSegmentator that are visible in the slice.
• multiple_components_same_label: Whether there are multiple disconnected components of the same anatomical structure in the slice. (We only asked questions about structures that appear once per slice.)
• question_answer
  • object_name: Anatomical names of the structures.
  • object_gray: Label index from the TotalSegmentator table (used for number markers; letter markers start at AA for 1 and continue with AB, AC, …).
  • object_center: x and y coordinates of the center of mass of the two structures the correct answers to the questions were derived from these centers).
  • question: The question presented to the model.
  • answer: The correct answer (0 = no, 1 = yes).
• rotate_flip_short: Information on how the image is roated and flipped (A: Flip: A1 0°, A2: 90°, A3: 180°, A4 270° || B: Not Flipped: B1 0°, B2: 90°, B3: 180°, B4 270°)

---

Fixed Prompt:

The json files contain the questions that we pass to the model.

In addition to the question, we add a fixed text prompt:

RQ1, RQ2, RQ3:

"The image is a 2D axial slice of an abdominal CT scan with soft tissue windowing. "
"Answer strictly with '1' for Yes or '0' for No. No explanations, no additional text. "
"Your output must contain exactly one character: '1' or '0'."
"Ignore anatomical correctness; focus solely on what the image shows.\n"
"Example:\n"
"Q: "Is the aorta above the spleen?" A: 1\n"
"Now answer the real question:\n\n"
f"Q: {question_from_json}"

AS:

"Answer strictly with '1' for Yes or '0' for No. No explanations, no additional text. "
"Your output must contain exactly one character: '1' or '0'."
"Focus solely on what the image shows.\n"
"Example:\n"
"Q: "Is the red dot above the blue dot" A: 1\n"
"Now answer the real question:\n\n"
f"Q: {question_from_json}"

The example questions in the fixed prompt are chosen to match the question in the task.
For example, RQ2 with dot marks: "Is the aorta (red) above the spleen (blue)?

The 2_inference_code/ already contains those fixed prompts.