Our study includes 16 biomedical image segmentation datasets. Place all the datasets in one directory under data to ease management. The file structure looks like

data/
├── <DATASET_NAME>/
│   ├── Prompts_Folder/
│   │   └── <prompt_files>        # text prompts (*.xlsx)
│   │
│   ├── Train_Folder/
│   │   ├── img/
│   │   │   └── <image_files>
│   │   └── label/
│   │       └── <mask_files>
│   │
│   ├── Val_Folder/
│   │   ├── img/
│   │   │   └── <image_files>
│   │   └── label/
│   │       └── <mask_files>
│   │
│   └── Test_Folder/
│       ├── img/
│       │   └── <image_files>
│       └── label/
│           └── <mask_files>
│
└── <DATASET_NAME_2>/
    └── (same structure as above)

Dataset Organization

Each dataset is split into training, validation, and testing splits.

The Prompts_Folder contains the text prompt files associated with each dataset. These include:

• Prompt definitions used for data-efficiency experiments (e.g., 10%, 25%, 50% training and validation subsets)
• Additional variant prompt designs explored in the study, such as alternative phrasing and semantic formulations

These prompt files enable flexible evaluation under different supervision regimes.

Dataset Summary

Dataset	Train	Validation	Test	Modality	Organ
BUSI	(62, 156, 312)	(7, 19, 39)	78	Ultrasound	Breast
BTMRI	(273, 684, 1,369)	(132, 330, 660)	1,005	MRI	Brain
ISIC	(80, 202, 404)	(9, 22, 45)	379	Dermatoscopy	Skin
Kvasir-SEG	(80, 200, 400)	(10, 25, 50)	100	Endoscopy	Colon
QaTa-COV19	(571, 1,429, 2,858)	(142, 357, 714)	2,113	X-ray	Chest
EUS	(2,631, 6,579, 13,159)	(175, 439, 879)	10,090	Ultrasound	Pancreas
BUSUC	567	122	122	Ultrasound	Breast
BUSBRA	1,311	282	282	Ultrasound	Breast
BUID	162	35	35	Ultrasound	Breast
UDIAT	113	25	25	Ultrasound	Breast
BRISC	4,000	1,000	1,000	MRI	Brain
UWaterlooSkinCancer	132	0	41	Dermatoscopy	Skin
CVC-ColonDB	20	0	360	Endoscopy	Colon
CVC-ClinicDB	490	61	61	Endoscopy	Colon
CVC-300	6	0	60	Endoscopy	Colon
BKAI	799	100	100	Endoscopy	Colon

Download the datasets

All the datasets can be found on Hugging Face here. Download each dataset seperately:

• BUSI Hugging Face
• BTMRI Hugging Face
• ISIC Hugging Face
• Kvasir-SEG Hugging Face
• QaTa-COV19 Hugging Face
• EUS Drive
• BUSUC Hugging Face
• BUSBRA Hugging Face
• BUID Hugging Face
• UDIAT Hugging Face
• BRISC Hugging Face
• UWaterlooSkinCancer Hugging Face
• CVC-ColonDB Hugging Face
• CVC-ClinicDB Hugging Face
• CVC-300 Hugging Face
• BKAI Hugging Face

After downloading each dataset, unzip and place each under data like the following

data/
├── BTMRI/
│   ├── Prompts_Folder/
│   │   └── <prompt_files>        # text prompts (*.xlsx)
│   │
│   ├── Train_Folder/
│   │   ├── img/
│   │   │   └── <image_files>
│   │   └── label/
│   │       └── <mask_files>
│   │
│   ├── Val_Folder/
│   │   ├── img/
│   │   │   └── <image_files>
│   │   └── label/
│   │       └── <mask_files>
│   │
│   └── Test_Folder/
│       ├── img/
│       │   └── <image_files>
│       └── label/
│           └── <mask_files>

Preprocessing EUS dataset

• Download the EUS Healthy subset, place it in data/EUS, extract the data, and rename the extracted content to EUS_healthy

• Download the EUS Cancer subset, place it in data/EUS, extract the data, and rename the extracted content to EUS_cancer

• Run the preprocessing script:

  python utils/preprocess_EUS.py
  

• Delete EUS_healthy and EUS_cancer