Abstract:
I present one of our recent studies related to a question of “how AI can adaptively optimize the measurement condition for phenotype discrimination”.
When the aim of a measurement such as classifying phenotypes is set, the conventional paradigm that separates data acquisition from analysis becomes inefficient. Our key lies in embedding AI-driven analysis directly into the measurement loop, dynamically guiding the process toward optimal acquisition conditions. We present our recent studies in AI-integrated measurement science, where artificial intelligence intervenes in real-time to accelerate data acquisition while preserving the accuracy of measurement objectives. When the aim of a measurement—such as detecting tumors, identifying anomalies, or classifying phenotypes—is defined, the conventional paradigm that separates data acquisition from analysis becomes increasingly inefficient. Our key component lies in embedding AI-driven analysis directly into the measurement loop, dynamically guiding the process toward optimal acquisition conditions.
This is particularly critical in applications constrained by time and cost, such as label-free phenotypic classification using spontaneous Raman microscopy. While Raman microscopy offers chemically rich information, its inherently slow acquisition—due to the weak scattering cross-section—limits practical throughput. We address this bottleneck by introducing a reinforcement learning-based illumination strategy, which adaptively determines where and how to measure during the imaging process. This method selectively focuses on regions likely to carry diagnostically relevant signals, drastically reducing acquisition time while maintaining discrimination accuracy [1].
[1] K. Tabata et al. On-the-fly Raman microscopy guaranteeing the accuracy of discrimination. PNAS 121(12), e2304866121, (2024)