Fig. 4: Computational detection of rare cells.

a–c Preliminary screening of the whole FOV to detect candidates for target cells (MCF7). At each scanning position, 120 frames of raw holograms were taken at 26.7 frames per second. Computational drift correction was applied to mitigate the horizontal shift caused by the fluid drift, where the vertical movement caused by the magnetic field was kept unmodified. The lateral position of each MCF7 candidate was located by CMA, maximum intensity projection and threshold-based detection. d–g Zoomed-in preliminary processing for the example region labelled ① in b, c. h–k Classification process for the two cell candidates labelled ① and ② in c. The axial location for each cell candidate was determined by autofocusing. A video was formed for each cell candidate by propagating each frame to the in-focus plane. The classification was performed by a densely connected P3D convolutional neural network, as detailed in the Methods section