From image stacks to tracked spines
During my internship in Cora Lai’s group at the University of Hong Kong, I worked on quantifying synaptic structural plasticity from in vivo two-photon images. I developed a pipeline that detects spines in 2D slices, links them across depth into 3D objects, and matches their identities across imaging sessions.
The workflow combines deep object detection, overlap-based linking across slices, and temporal matching using motion, depth, and appearance. This turns manual image-by-image comparisons into structured tracks for studying spine formation and elimination.
2D detection
The detector achieved F1 = 0.77 ± 0.03 in the project evaluation, compared with 0.68 ± 0.04 for the Vogel baseline and 0.39 ± 0.07 for DeepD3. The montage below shows example detections on individual optical slices.
3D association across depth
Linking detections across neighbouring slices produced more consistent 3D spine identities. HOTA was 98.10 ± 0.34, compared with 90.43 ± 1.57 for the connected-component baseline.
4D tracking across sessions
Across time, the pipeline achieved HOTA = 88.37 and IDF1 = 82.81. The example shows how identities can be matched between day 1 and day 2 despite differences in position and imaging depth.
Outcome
I delivered the analysis pipeline to laboratory members to support routine image annotation and analysis.