GALS-CE: A Generative AI Model for Liver Lesion Screening Integrating Contrast Agent Knowledge

GALS-CE is an innovative medical imaging AI system developed by the SMU Medical Vision team. Its core lies in integrating generative artificial intelligence and contrast agent kinetics knowledge to enable intelligent screening and classification of liver lesions using multi-phase CT images. This model aims to address the problems of traditional CT diagnosis relying on physicians' experience and insufficient integration of medical prior knowledge in deep learning, providing a new solution for early screening and precise diagnosis of liver lesions.

Early screening and precise diagnosis of liver lesions are important topics in clinical medicine. Traditional CT image diagnosis is highly dependent on radiologists' experience and requires analysis of multi-phase image features (especially contrast agent distribution patterns). Deep learning has great potential in the field of medical imaging, but how to effectively integrate medical prior knowledge (such as contrast agent kinetics) with data-driven methods remains an open question, and GALS-CE was proposed in this context.

GALS-CE is a two-stage deep learning framework, with its core innovation being the integration of generative AI and contrast agent knowledge. Two-stage training paradigm: In the first stage, the generative model learns multi-phase CT mapping and can synthesize subsequent phase images; in the second stage, the classification network is trained using synthetic and real images to achieve lesion screening and benign/malignant classification. The contrast agent knowledge fusion mechanism explicitly models its dynamic distribution rules, ensuring that synthetic images comply with medical physics laws and improving classification reliability.

Developed based on Python 3.8, it relies on PyTorch 2.0 (CUDA 11.8) and medical imaging processing libraries. It is recommended to use a conda isolated environment to ensure reproducibility. Data organization specifications: Each case contains multi-phase images in NifTI format (NC, AP, PVP, DP) and optional mask files. The standardized structure facilitates batch processing and multi-center collaboration. The training process supports quick testing and full training (custom hyperparameters, specified GPU). After training, it automatically performs inference to generate results and visualizations, and also supports an independent inference mode.

1. Improve diagnostic consistency: Reduce subjective differences in physicians' interpretation of phase features and provide an objective and reproducible auxiliary tool; 2. Alleviate data scarcity: The generative model can synthesize complete enhanced sequences to facilitate small-sample learning; 3. Ensure image integrity: Infer missing images based on existing phases to assist physicians in comprehensive evaluation (e.g., when patients cannot complete all phase scans).

Current constraints: Only for liver CT images; generalization to other organs/modalities needs verification; the impact of subtle differences between synthetic quality and real images requires large-scale clinical validation. Future directions: Explore interpretability technologies (attention visualization, saliency maps) to enhance physician trust; integrate more clinical information (laboratory indicators, medical history, genomic data) to build a multi-modal system; use GAN or diffusion models to improve image synthesis quality.

GALS-CE represents the knowledge-driven direction of medical imaging AI, encoding contrast agent kinetics into neural networks and demonstrating the potential of knowledge-driven AI in precision medicine. With the maturity of technology and in-depth clinical validation, it is expected to become a powerful assistant for radiologists and benefit more patients.