💻 computer science

CPS4: Class Prompt driven Semi-Supervised Spine Segmentation with Class-specific Consistency Constraint

The paper proposes CPS4, a novel text-guided semi-supervised spine segmentation framework that leverages class-specific consistency constraints during VLM pretraining to generate high-quality pseudo labels, achieving superior performance with only 5% labeled data.

Original authors: Qingtao Pan, Hongzan Sun, Bing Ji, Shuo Li

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Qingtao Pan, Hongzan Sun, Bing Ji, Shuo Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to identify and outline every single vertebra and disc in a human spine from an MRI scan. This is a tough job because the spine has many parts that look very similar, and labeling every single pixel in a medical image is like counting grains of sand on a beach—it takes forever and requires a highly trained expert.

This paper introduces a new method called CPS4 to solve this problem. It's designed to work even when the robot only has a few "labeled" examples (where a human has already drawn the outlines) and a huge pile of "unlabeled" images (where no one has drawn anything yet).

Here is how CPS4 works, broken down into simple concepts:

The Problem: The Robot Gets Confused

Usually, when robots try to learn from unlabeled images, they guess the outlines (called "pseudo-labels") and then try to learn from their own guesses. The problem is, if the robot makes a small mistake early on, it keeps making that mistake over and over, getting worse and worse. It's like a student trying to learn math by copying answers from a friend who doesn't know the answers either.

The Solution: A "Text Guide"

The authors realized that while the robot is bad at guessing, it's actually pretty good at understanding language. They decided to use text prompts (like writing "This is the Lumbar Vertebra 1") to help the robot focus.

Think of the robot as a very smart but slightly distracted artist. If you just say, "Draw a spine," the artist might draw a messy blob. But if you hand the artist a specific instruction card that says, "Focus only on the Lumbar Vertebra 1," the artist can draw that specific part much more accurately.

How CPS4 Works (The Two-Step Dance)

The method uses a two-stage training process, which the authors call CPS4:

Stage 1: The "Strict Teacher" Phase (Pretraining)
Before the robot starts guessing on unlabeled images, it needs to learn how to listen to the text instructions perfectly.

The Analogy: Imagine a teacher showing a student a picture of a specific bone and saying, "This is the T12 vertebra." The student must learn to point exactly at that bone and ignore everything else.
The Innovation: The authors created two special "rules" (loss functions) to force the robot to pay attention:
1. Token-Level Attention: This ensures the robot knows which word in the sentence corresponds to which part of the image. It's like making sure the word "T12" is glued to the T12 bone in the robot's mind.
2. Pixel-Level Attention: This ensures the robot doesn't just point vaguely at the bone but covers the entire bone accurately, not just a tiny corner of it.
The Result: The robot learns to tightly couple the text description with the actual image part.

Stage 2: The "Helper" Phase (Semi-Supervised Segmentation)
Now, the robot is ready to tackle the unlabeled images.

The Process: For every unlabeled spine image, the robot uses its trained "text guide" to generate a separate map for each type of spine part (e.g., one map for T10, one for T11, etc.).
The Magic: It combines all these individual maps into one big, high-quality "master map." This master map is much better than a guess because it was guided by the text prompts.
The Loop: The robot uses this high-quality "text-guided map" to teach itself, correcting its own mistakes and learning faster than it could on its own.

The Results: A Big Win with Little Data

The researchers tested this on a public spine dataset.

The Challenge: They tried to train the robot using only 5% of the labeled data (a tiny amount).
The Outcome: Even with so little data, CPS4 achieved a Dice score of 80.44%.
Comparison: This was better than all other popular methods, including those that use text (Vision-Language Models) and those that don't. It proved that using text prompts to guide the robot makes the "guesses" much more accurate.

Visual Proof

The paper shows pictures (Figures 3 and 5) comparing their method to others.

Other methods: Often get confused, mixing up different vertebrae or missing parts entirely.
CPS4: The "attention maps" (the robot's internal focus) show that it knows exactly where to look. When the text says "L5," the robot's focus is perfectly locked onto the L5 vertebra, ignoring the rest of the spine.

Summary

In short, CPS4 is a system that teaches a computer to segment spines by using text descriptions as a spotlight. By first training the computer to understand exactly which text belongs to which bone, and then using that understanding to generate better practice examples for itself, the system can learn to map spines very accurately even when humans haven't provided many examples to start with.

Technical Summary: CPS4 – Class Prompt driven Semi-Supervised Spine Segmentation with Class-specific Consistency Constraint

Problem Statement

Automatic spine segmentation, specifically the multi-class segmentation of vertebral bodies (VBs) and intervertebral discs (IVDs) in MRI, is critical for orthopedic diagnosis and surgical planning. While deep learning methods have advanced significantly, they rely heavily on pixel-level annotations, which are labor-intensive and require expert knowledge. Semi-supervised learning (SSL) offers a solution by utilizing limited labeled data alongside abundant unlabeled data. However, existing SSL methods, which typically rely on consistency regularization and pseudo-labeling, often suffer from low-quality pseudo-labels. These noisy labels introduce incorrect supervision, causing models to focus on irrelevant features and hindering the effective exploitation of unlabeled data.

Although Vision-Language Models (VLMs) show potential for enhancing pseudo-label quality by leveraging textual class prompts to generate segmentation maps, current VLMs lack explicit constraints to ensure consistency between class prompts and specific image objects. This deficiency leads to weak class-specific semantic binding and cross-class interference, resulting in unsatisfactory performance in multi-class segmentation map generation.

Methodology

The authors propose CPS4, a novel framework that integrates class prompts into semi-supervised spine segmentation through a two-stage training strategy.

Stage 1: Class-Specific Consistency Constrained VLM Pretraining

Existing VLMs (e.g., CLIP, MedCLIP) perform cross-modal alignment at the image level, which fails to ensure that distinct class prompts focus on their corresponding objects. To address this, CPS4 introduces a pretraining stage using a Vision-Language Model (specifically utilizing ViT for vision and BioClinicalBERT for text) with two novel loss functions to enforce consistency between class prompts and spine units:

Token-Level Attention Loss ( $L_{token}$ ): This loss optimizes the alignment between a class prompt token ( $F_{T,i}$ ) and the corresponding image patch tokens ( $F_I$ ). It calculates a similarity score to generate a coarse attention map ( $A_i$ ) and aggregates activations toward the predicted spine unit region ( $B_i$ ). The goal is to force the textual prompt to couple closely with the target spine unit in the semantic space.
Pixel-Level Attention Loss ( $L_{pixel}$ ): To prevent the model from over-aggregating attention into only certain subregions of a target unit, a pixel-level cross-entropy objective is added. This ensures the attention map covers the entire target spine unit uniformly.

The total pretraining loss is the average of $L_{token}$ and $L_{pixel}$ .

Stage 2: Class Prompt Driven Semi-Supervised Spine Segmentation

In this stage, the pretrained vision-text encoders are used to enhance the pseudo-labeling process for unlabeled data.

Class Prompt-Guided Segmentation Map Generation: For each unlabeled spine image, the system generates a binary segmentation map for every class (e.g., "Lumbar Vertebra 1") by computing the dot product between image patch tokens and class prompt tokens. These maps are upsampled and converted to binary masks using the OTSU automatic threshold method. All class-specific maps are stacked to form a unified multi-class segmentation map ( $y^{text}$ ).
Pseudo Label Generation & Optimization: The framework employs a Mean Teacher (MT) architecture with a student and a teacher network. The teacher network generates pseudo-labels ( $y^t_u$ $y_{u}^{t}$ ) for unlabeled images via Exponential Moving Average (EMA).
- Supervised Loss ( $L_{sup}$ ): Calculated between student predictions on labeled data and ground truth.
- Semi-Supervised Loss ( $L_{semi}$ ): Composed of two parts:
  1. $L^{text}_{semi}$ : Consistency loss between the student's prediction on unlabeled data and the class prompt-generated map ( $y^{text}$ ).
  2. $L^{enhance}_{semi}$ : Consistency loss between the student's prediction and an enhanced pseudo-label created by integrating the teacher's pseudo-label ( $y^t_u$ ) and the class prompt map ( $y^{text}$ ).

Key Contributions

First Text-Guided SSL for Spine Segmentation: This work is the first to introduce textual class prompts to enhance pseudo-label quality in semi-supervised spine segmentation.
Novel Class-Specific Consistency Constraint: The authors propose a specific pretraining mechanism using token- and pixel-level attention losses to enforce explicit constraints, ensuring tight semantic coupling between each class prompt and its target spine unit.
State-of-the-Art Performance: The method demonstrates superior performance across various labeled data ratios, particularly in low-data regimes.

Experimental Results

The method was evaluated on the MRSpineSeg dataset, containing T2-weighted MR volumetric images of 215 subjects with 20 categories (19 spinal structures and background).

Performance: Using only 5% labeled data, CPS4 achieved a Dice score of 80.44% and an mIoU of 70.63%.
Comparison: These results surpass popular semi-supervised learning methods (e.g., MT, BCP, SS-Net) and existing VLM-based approaches (e.g., CLIP, MedCLIP, MGCA, EGMA).
- Compared to the suboptimal GraphCL method with text prompts on 5% labeled data, CPS4 improved mDice by 2.23% and mIoU by 2.42%.
Robustness: Ablation studies confirmed that both the token-level and pixel-level losses contribute significantly to performance. The method showed stable performance gains as the ratio of labeled data increased from 5% to 50%, outperforming the MT baseline by a clear margin.
Visualization: Attention maps generated by CPS4 showed strong alignment with target spine units, whereas baseline VLMs (like MedCLIP) exhibited scattered activations and weak semantic binding.

Significance

The paper claims that CPS4 effectively mitigates the label scarcity problem in spine segmentation by leveraging the semantic power of text to guide the generation of high-quality pseudo-labels. By explicitly constraining the consistency between class prompts and spine units during pretraining, the framework overcomes the limitations of current VLMs in medical imaging tasks. The results suggest that this approach enables more reliable class prompt-driven pseudo-label generation, significantly improving the efficacy of semi-supervised learning in scenarios with limited annotated data. The authors note that their code will be made available to facilitate further research.