MedCLIPSeg: Probabilistic Vision–Language Adaptation for Data-Efficient and Generalizable Medical Image Segmentation

Overview

Medical image segmentation remains challenging due to limited annotations for training, ambiguous anatomical features, and domain shifts. While vision–language models such as CLIP offer strong cross-modal representations, their potential for dense, text-guided medical image segmentation remains underexplored. We present MedCLIPSeg, a novel framework that adapts CLIP for robust, data-efficient, and uncertainty-aware medical image segmentation. Our approach leverages patch-level CLIP embeddings through probabilistic cross-modal attention, enabling bidirectional interaction between image and text tokens and explicit modeling of predictive uncertainty. Together with a soft patch-level contrastive loss that encourages nuanced semantic learning across diverse textual prompts, MedCLIPSeg improves data efficiency and domain generalizability. Extensive experiments across 16 datasets, spanning five imaging modalities and six organs, demonstrate that MedCLIPSeg outperforms prior methods in accuracy, efficiency, and robustness, while providing interpretable uncertainty maps that highlight the local reliability of segmentation results. This work demonstrates the potential of probabilistic vision–language modeling for text-driven medical image segmentation.

How to install datasets

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>

Citation

If you use our work, please consider citing:

@inproceedings{koleilat2026medclipseg,
  title     = {MedCLIPSeg: Probabilistic Vision--Language Adaptation for Data-Efficient and Generalizable Medical Image Segmentation},
  author    = {Koleilat, Taha and Asgariandehkordi, Hojat, Nejatimanzari, Omid and Barile, Berardino and Xiao, Yiming and Rivaz, Hassan},
  booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference},
  year      = {2026}
}