Thorny but rosy: prosperities and difficulties in ‘AI plus medicine’ concerning data collection, model construction and clinical deployment

[...]we discussed the differences between mental health and other domains as well as the special considerations in ‘AI plus psychiatry’. The framework maintains the excellent performance of the AI model for pneumonia classification and liver tumour segmentation while ensuring strict privacy protection. [...]federated learning (FL) makes cross-centre data communication possible. [...]the models tend to suffer from overfitting to specific datasets, resulting in a lack of robustness and generalisability when applied to external validation or prospective cohorts. Heatmaps are commonly used to visualise the attention regions of a deep learning model, helping researchers better understand the model’s decision-making process. [...]integrating medical prior knowledge can be another approach to improve the reliability of the AI models.