Physics-based machine learning for subcellular segmentation in living cells

Segmenting subcellular structures in living cells from fluorescence microscope images is a ground truth (GT)-deficient problem. The microscopes’ three-dimensional blurring function, finite optical resolution due to light diffraction, finite pixel resolution and the complex morphological manifestations of the structures all contribute to GT-hardness. Unsupervised segmentation approaches are quite inaccurate. Therefore, manual segmentation relying on heuristics and experience remains the preferred approach. However, this process is tedious, given the countless structures present inside a single cell, and generating analytics across a large population of cells or performing advanced artificial intelligence tasks such as tracking are greatly limited. Here we bring modelling and deep learning to a nexus for solving this GT-hard problem, improving both the accuracy and speed of subcellular segmentation. We introduce a simulation-supervision approach empowered by physics-based GT, which presents two advantages. First, the physics-based GT resolves the GT-hardness. Second, computational modelling of all the relevant physical aspects assists the deep learning models in learning to compensate, to a great extent, for the limitations of physics and the instrument. We show extensive results on the segmentation of small vesicles and mitochondria in diverse and independent living- and fixed-cell datasets. We demonstrate the adaptability of the approach across diverse microscopes through transfer learning, and illustrate biologically relevant applications of automated analytics and motion analysis.

Advancing COVID-19 diagnosis with privacy-preserving collaboration in artificial intelligence

Artificial intelligence provides a promising solution for streamlining COVID-19 diagnoses; however, concerns surrounding security and trustworthiness impede the collection of large-scale representative medical data, posing a considerable challenge for training a well-generalized model in clinical practices. To address this, we launch the Unified CT-COVID AI Diagnostic Initiative (UCADI), where the artificial intelligence (AI) model can be distributedly trained and independently executed at each host institution under a federated learning framework without data sharing. Here we show that our federated learning framework model considerably outperformed all of the local models (with a test sensitivity/specificity of 0.973/0.951 in China and 0.730/0.942 in the United Kingdom), achieving comparable performance with a panel of professional radiologists. We further evaluated the model on the hold-out (collected from another two hospitals without the federated learning framework) and heterogeneous (acquired with contrast materials) data, provided visual explanations for decisions made by the model, and analysed the trade-offs between the model performance and the communication costs in the federated training process. Our study is based on 9,573 chest computed tomography scans from 3,336 patients collected from 23 hospitals located in China and the United Kingdom. Collectively, our work advanced the prospects of utilizing federated learning for privacy-preserving AI in digital health.