Interactive Model Training for Multi-modal Spatial Biology using Human Feedback

Multiscale spatial biology technologies, including volume electron microscopy, super-resolution imaging, and spatial transcriptomics, allow cells and their molecular states to be studied in situ within their natural environments. These approaches generate large, complex, multi-modal datasets, but their analysis using AI remains limited by the need for extensive and often subjective manual annotation. Producing accurate and detailed labels for tasks such as cell segmentation or aligning image features to gene expression is time-consuming, inconsistent across experts, and difficult to scale across experiments and modalities.

This project aims to develop advanced AI methods that enable models to be trained using small amounts of structured human feedback instead of dense annotations. Rather than requiring experts to label data exhaustively, models will be trained through simple feedback signals, such as preferences between alternative outputs, quality ratings, or binary judgements, integrated directly into the learning process using reinforcement learning and related optimisation techniques. These methods will be combined with few-shot and meta-learning strategies to enable rapid model adaptation from limited expert interaction.

The approach will be evaluated on challenging multi-modal spatial biology tasks, including cell segmentation in volume EM and super-resolution microscopy, and learning relationships between spatial image features and omics measurements such as gene expression profiles. By reducing reliance on manual annotation while improving robustness and transferability, the project aims to deliver AI methods that can be trained with reduced expert workload, and can efficiently adapt to new spatial datasets, modalities and studies.