LST-AI is a deep learning-based tool for accurate segmentation of white matter (WM) lesions in multiple sclerosis (MS). It is an advanced extension of the previously developed LST tool, which was an engineered lesion segmentation tool. LST-AI uses an ensemble of three 3D U-Nets to improve segmentation of heterogeneous MS lesions. It addresses the imbalance between WM lesions and non-lesioned WM by using a composite loss function that combines binary cross-entropy and Tversky loss. The tool was trained on 491 pairs of T1-weighted and FLAIR images from a 3T MRI scanner, with expert neuroradiologists manually segmenting the lesion maps. LST-AI also includes a lesion location annotation tool that labels lesions according to the 2017 McDonald criteria. It was evaluated on 103 test cases using publicly available data and compared with several publicly available lesion segmentation models. The empirical analysis showed that LST-AI achieved superior performance compared to existing methods, with Dice and F1 scores exceeding 0.62, outperforming LST, SAMSEG, and the nnUNet framework. LST-AI demonstrated exceptional performance on the MSSEG-1 challenge dataset, with a Dice score of 0.65 and an F1 score of 0.63. With increasing lesion volume, the lesion detection rate rapidly increased, reaching over 75% for lesions between 10 and 100 mm³. LST-AI is recommended for research groups currently using LST, as it is an open-source tool available as a command-line tool, dockerized container, or Python script. The tool is designed for research and non-clinical purposes and has not undergone clinical certification or licensing. It is intended for a diverse user base and can be used in similar ways as Free-surfer/FSL command line tools or nicMSlesions (docker). LST-AI is recommended for use in GPU-enabled environments but also provides a fallback method for CPU-only usage. The tool is available for three different workflows: segmentation-only, lesion location annotation-only, or both. Labels can be exported in the original subject space or in the MNI ICBM152 template space. The source code is available for the community to adapt and tailor the tools for different application scenarios. The tool is continuously maintained and updated in the GitHub repository. LST-AI is evaluated on multiple external datasets, showing that it generalizes well to data from different centers and scanners without retraining. It outperforms benchmark methods in all categories except PPV and PPVL. LST-AI achieves higher DSC and F1 scores compared to other methods, indicating superior segmentation performance on both voxel-wise and lesion-wise levels. The lowest ASD is also obtained with LST-AI, indicating more accurate lesion contouring compared to the benchmark methods. Overall, the results show thatLST-AI is a deep learning-based tool for accurate segmentation of white matter (WM) lesions in multiple sclerosis (MS). It is an advanced extension of the previously developed LST tool, which was an engineered lesion segmentation tool. LST-AI uses an ensemble of three 3D U-Nets to improve segmentation of heterogeneous MS lesions. It addresses the imbalance between WM lesions and non-lesioned WM by using a composite loss function that combines binary cross-entropy and Tversky loss. The tool was trained on 491 pairs of T1-weighted and FLAIR images from a 3T MRI scanner, with expert neuroradiologists manually segmenting the lesion maps. LST-AI also includes a lesion location annotation tool that labels lesions according to the 2017 McDonald criteria. It was evaluated on 103 test cases using publicly available data and compared with several publicly available lesion segmentation models. The empirical analysis showed that LST-AI achieved superior performance compared to existing methods, with Dice and F1 scores exceeding 0.62, outperforming LST, SAMSEG, and the nnUNet framework. LST-AI demonstrated exceptional performance on the MSSEG-1 challenge dataset, with a Dice score of 0.65 and an F1 score of 0.63. With increasing lesion volume, the lesion detection rate rapidly increased, reaching over 75% for lesions between 10 and 100 mm³. LST-AI is recommended for research groups currently using LST, as it is an open-source tool available as a command-line tool, dockerized container, or Python script. The tool is designed for research and non-clinical purposes and has not undergone clinical certification or licensing. It is intended for a diverse user base and can be used in similar ways as Free-surfer/FSL command line tools or nicMSlesions (docker). LST-AI is recommended for use in GPU-enabled environments but also provides a fallback method for CPU-only usage. The tool is available for three different workflows: segmentation-only, lesion location annotation-only, or both. Labels can be exported in the original subject space or in the MNI ICBM152 template space. The source code is available for the community to adapt and tailor the tools for different application scenarios. The tool is continuously maintained and updated in the GitHub repository. LST-AI is evaluated on multiple external datasets, showing that it generalizes well to data from different centers and scanners without retraining. It outperforms benchmark methods in all categories except PPV and PPVL. LST-AI achieves higher DSC and F1 scores compared to other methods, indicating superior segmentation performance on both voxel-wise and lesion-wise levels. The lowest ASD is also obtained with LST-AI, indicating more accurate lesion contouring compared to the benchmark methods. Overall, the results show that