TLDR: LangMamba is a new framework that uses vision-language models (VLMs) to improve low-dose CT (LDCT) image denoising. It pre-trains a Language-guided AutoEncoder (LangAE) to understand anatomical semantics, then uses a Semantic-Enhanced Efficient Denoiser (SEED) and a Language-engaged Dual-space Alignment (LangDA) Loss to produce high-quality, semantically aligned denoised images. The method outperforms existing techniques, offers better detail preservation and visual fidelity, and provides language-level explainability, all while being computationally efficient.
Low-dose computed tomography (LDCT) is a valuable medical imaging technique that reduces radiation exposure for patients. However, this benefit often comes at the cost of degraded image quality, which can make it harder for doctors to make accurate diagnoses. Traditional deep learning methods for denoising these images primarily focus on pixel-level improvements, often missing the bigger picture of high-level semantic information that could guide better image reconstruction.
Recent advancements in vision-language models (VLMs), which combine visual and textual understanding, have opened new doors. These models can capture structured semantic information using language, offering a fresh approach to improving LDCT image quality. This is where LangMamba comes in, a new framework designed to leverage these VLM-derived representations to enhance the denoising process, guided by normal-dose CT (NDCT) images.
LangMamba operates through a clever two-stage learning strategy. The first stage involves pre-training a “Language-guided AutoEncoder” (LangAE). This LangAE uses pre-trained VLMs to map NDCT images into a semantic space that is rich with anatomical information. Think of it as teaching the system to understand the “meaning” of different body parts in the CT scan.
In the second stage, LangAE works in synergy with two crucial components to guide the LDCT denoising. The first is the “Semantic-Enhanced Efficient Denoiser” (SEED). SEED enhances the local semantic information relevant to NDCT images while also capturing global features efficiently using a mechanism inspired by the Mamba architecture, known for its ability to process long sequences of data effectively. The second component is the “Language-engaged Dual-space Alignment” (LangDA) Loss. This innovative loss function ensures that the denoised images align closely with the normal-dose CT images in both their visual appearance (perceptual space) and their underlying semantic meaning (semantic space).
A key advantage of LangMamba is that the large-scale VLMs are not directly involved in the training or inference of the denoising process itself. This significantly reduces the computational burden and makes the system more practical for clinical use. The LangAE, once trained, can even generalize well to new, unseen datasets, which helps reduce training costs and deployment time in new scenarios.
Extensive experiments conducted on two public datasets have shown that LangMamba outperforms existing state-of-the-art methods. It significantly improves the preservation of fine details and enhances the overall visual fidelity of the CT images. Furthermore, the LangDA loss contributes to better explainability by integrating language-guided insights into the image reconstruction, making the results more understandable and trustworthy for radiologists. This work represents a pioneering effort in applying VLMs for LDCT denoising, highlighting the immense potential of language as a supervisory signal to advance medical image processing.
Also Read:
- PathCoT: Enhancing AI’s Understanding of Pathology Images with Expert Reasoning
- MML-SurgAdapt: A Unified AI Framework for Multi-Task Surgical Vision with Reduced Labeling
For more technical details, you can refer to the original research paper: LangMamba: A Language-driven Mamba Framework for Low-dose CT Denoising with Vision-language Models.
