Search papers, labs, and topics across Lattice.
This study introduces a secure multi-party computation (MPC) framework that enables the training and inference of the CellCnn convolutional neural network on sensitive single-cell data without compromising patient privacy. By allowing multiple institutions to collaborate on rare disease detection while keeping their data confidential, the method addresses the challenge of limited patient cohorts in studying diseases like leukaemia and viral infections. The approach not only maintains accuracy comparable to traditional methods but also surpasses previous privacy-preserving techniques by retaining critical components of the CNN architecture.
Achieving high-accuracy disease detection without ever exposing sensitive patient data could revolutionize collaborative medical research.
The detection of rare disease-associated cell subsets from high-dimensional single-cell measurements is critical for understanding diseases such as leukaemia and viral infections. CellCnn, a convolutional neural network (CNN) designed for this task, has demonstrated the ability to identify phenotype-associated cell populations at frequencies as low as 0.01\%. Training such models reliably requires patient cohorts that are larger and more diverse than any single institution can typically assemble, and the underlying single-cell data is too sensitive to share across institutional boundaries under existing privacy regulations. We propose a secure multi-party computation (MPC) framework that enables the training and inference of CellCnn entirely on secret-shared data. This ensures that neither the participants nor the computing servers ever observe raw patient data or intermediate values. Evaluated on benchmark single-cell datasets for cytomegalovirus infection (CMV) and acute myeloid leukaemia (AML), our implementation preserves accuracy close to its plaintext counterpart while outperforming the prior privacy-preserving baseline. In contrast to earlier privacy-preserving approaches that removed components such as ReLU activations and bias terms, our method retains these key parts of the CellCnn architecture and supports accurate analysis without exposing raw patient data.
