Artificial intelligence driven multimodal medical image segmentation algorithm: applied research and clinical verification

Abstract Multimodal medical imaging—comprising CT, MRI, and PET—is fundamental to modern diagnosis, lesion localization, and treatment planning. The accuracy and speed of segmenting these modalities directly affect clinical decisions. To overcome limitations in cross-modal integration, sensitivity to small lesions, generalization, and workflow compatibility, we propose AMIS-Net, an AI-driven segmentation network. Built on an encoder-decoder backbone, AMIS-Net incorporates a Dual Attention Module (DAM) for adaptive feature recalibration and a Small Object Capture (SOC) module for multi-scale feature extraction. A hybrid loss function balances pixel-wise training to address severe class imbalance. Extensive validation on CHAOS, Synapse, and a proprietary clinical dataset shows that AMIS-Net consistently outperforms U-Net, ResUNet, and STUNet. On Synapse, it achieves a Dice of 83.17% and HD95 of 20.89 mm, with per-organ Dice scores ranging from 74.85% (esophagus) to 94.21% (liver). Deployed in a clinical system for liver tumors and intracranial hemorrhage, AMIS-Net reduces median reading time from 8.5 to 4.2 min for senior radiologists and from 12.3 to 5.7 min for junior radiologists, improves diagnostic accuracy, and lowers missed diagnosis rates. This work offers a robust, clinically deployable solution for multimodal image analysis and advances the integration of AI into biomedical practice.