# Explainable Radiologist-Aligned VLM for CT Image Quality Assessment

**Authors:** **Jiajun Wang**, Yipeng Sun, Siming Bayer, Andreas Maier  
**Affiliation:** Pattern Recognition Lab, Friedrich-Alexander Universität Erlangen-Nürnberg, Germany  
**Links:** [GitHub Repository](https://github.com/atJesse/VLM-CT-IQA)

* * *

## 📖 Abstract

The assessment of computed tomography (CT) image quality has traditionally relied on manual evaluation by radiologists—a method that is both subjective and time-consuming. While Deep Learning methods exist, they often only give quantitative scores and lack explainability. To address this, we propose a **parameter-efficient supervised fine-tuning (SFT) framework** for the medical VLM, **MedGemma-4B-IT**. By employing **Quantized Low-Rank Adaptation (QLoRA)**, we aligned the model's visual perception with expert quantitative judgment.

Our results demonstrate a substantial improvement in correlation with expert scores (**SRCC=0.7950**, **PLCC=0.7907**) , significantly outperforming zero-shot baselines like Gemini 2.5 Pro and Gemini 2.5 Flash. Most importantly, our model generates **professional textual explanations** that emulate the reasoning and explanation style of radiologists.

* * *

## 💡 Motivation: Why Explainable AI for CT?

Diagnostic utility in CT scans depends critically on image quality. Degradation due to noise, artifacts, or insufficient contrast can lead to misdiagnoses and repeated examinations.

*   **The Problem with Manual Scoring:** It is labor-intensive, time-consuming, and prone to inter-observer variability.
    
*   **The Problem with Previous AI:** Conventional Deep Learning methods (NR-IQA) provide a score but remain opaque, making it difficult to interpret why specific scores are assigned.
    
*   **The Problem with Cloud VLMs:** General-purpose, closed-source VLMs are often constrained by patient privacy regulations.
    

Our goal was to create a **locally deployable, privacy-preserving, and explainable** solution.

* * *

## 🛠️ Methodology: The Framework

We formulated CT-IQA (Image Quality Assessment) as a multimodal reasoning task.

### 1\. The Model Architecture

We selected **MedGemma-4B-IT** as our base model due to its strong medical priors. The architecture consists of:

*   **Vision Encoder:** SigLIP, to capture local anatomical and noise patterns.
    
*   **Multimodal Projector:** Aligns visual representations with the language space.
    
*   **Language Model:** Gemma, generating both the textual reasoning and the final quality scores.
    
    ![](https://cdn.hashnode.com/uploads/covers/698c94f239413f8a70063736/1bb83288-e999-4180-b389-27af9790c989.png align="center")
    

### 2\. Parameter-Efficient Fine-Tuning (QLoRA)

To make training efficient, we used **QLoRA**.

*   **4-bit Quantization:** The backbone weights are quantized to 4-bit precision and frozen to reduce memory consumption.
    
*   **Trainable Adapters:** Only the low-rank adapters (~1% of parameters) are trainable.
    

### 3\. Data Construction with "Teacher" Distillation

Since large labeled datasets with explanations are scarce, we created a novel pipeline using the **LDCTIQA dataset** (1,000 CT slices):

*   **Teacher Model:** We employed **Gemini 2.5 Pro** (the best zero-shot performer) to generate expert-level textual explanations for the training data.
    
*   **Fine-tuning:** We trained our model to mimic these high-quality explanations, pairing them with radiologist scores.
    

* * *

## 📊 Results

### Quantitative Performance

Our fine-tuned model achieved state-of-the-art results compared to zero-shot baselines on the test set.

| Model | SRCC (Correlation) | PLCC (Linearity) | MAE (Error) |
| --- | --- | --- | --- |
| **MedGemma-4B-IT (Fine-tuned)** | **0.7950** | **0.7907** | **0.5780** |
| Gemini 2.5 Pro (Zero-shot) | 0.7328 | 0.7204 | 0.6540 |
| Gemini 2.5 Flash (Zero-shot) | 0.7170 | 0.6946 | 0.7360 |
| MedGemma-4B-IT (Zero-shot) | \-0.2438 | \-0.2029 | 1.4790 |

*Data Source: Table 1 in Paper*

The fine-tuning improved the SRCC by over **+1.0** compared to the original weights.

### Qualitative Analysis (Case Study)

In testing, when presented with a low-quality image (Ground Truth Score: 1.2), the zero-shot baseline incorrectly rated it as high quality (3.2).

![](https://cdn.hashnode.com/res/hashnode/image/upload/v1770903965417/0e79b429-3500-4b2f-bb02-970a3d01745f.png align="center")

**Our Fine-Tuned Model:**

*   **Predicted Score:** 1.0 (Very close to GT 1.2).
    
*   **Generated Reasoning:** Correctly identified "Severe Artifacts," "Streak Artifacts," and "High Noise". It explicitly noted that streaks were radiating from the pelvic girdle, obscuring tissue texture.
    

* * *

## 🚀 Conclusion

We demonstrated that a specialized medical VLM can be fine-tuned to emulate the reasoning style of radiologists. This provides a **locally deployable, privacy-preserving, and explainable** tool for automated CT image quality assessment.

### 🔗 Resources

*   **Code:** [GitHub - VLM-CT-IQA](https://github.com/atJesse/VLM-CT-IQA)
    

* * *

### 📝 Citation

If you find this work helpful, please consider citing our paper:

```plaintext
@inproceedings{wang2025explainable,
  title={Explainable Radiologist-Aligned VLM for CT Image Quality Assessment},
  author={Wang, Jiajun and Sun, Yipeng and Bayer, Siming and Maier, Andreas},
  booktitle={German Conference on Medical Image Computing (BVM)},
  year={2026}
}
```