Featured 2024 ML / research engineer Solo (research project) Placeholder — e.g. 4 months

Gap Junction Segmentation in Electron Microscopy

Deep-learning segmentation pipeline for detecting thin gap-junction structures in electron-microscopy volumes.

GitHub

TL;DR

An end-to-end pipeline that segments extremely thin gap-junction membranes in large electron-microscopy (EM) volumes, handling severe class imbalance and comparing 2D and 3D model designs. (Placeholder description — replace with your real project narrative.)

Problem

Gap junctions are sub-pixel-thin structures that are easy to miss and hard to label, with a tiny fraction of foreground pixels per slice. Standard segmentation models collapse to predicting all-background. (Placeholder — refine with specifics.)

Note: This is placeholder case-study content. Replace the prose below with your real write-up. The structured sections (Methods, Results, etc.) are generated from the frontmatter above.

Why it matters

Mapping gap junctions in electron microscopy supports connectomics and neuroscience research. Because the structures are so thin, manual annotation is slow and error-prone, so a reliable automated segmentation pipeline saves significant expert time.

Dataset / inputs

EM image volumes with sparse, expert-provided gap-junction masks. (Placeholder — describe resolution, volume size, and how masks were created. Do not include any private or proprietary data.)

Technical decisions

I started with a 2D U-Net baseline to establish a fast iteration loop, then introduced a 3D patch-based model to capture context across slices. The biggest lever was the loss function: with such extreme foreground/background imbalance, a combined Dice + focal loss outperformed plain cross-entropy. (Placeholder — swap in your real findings.)

Challenges

Severe class imbalance, limited labeled data, and the memory cost of 3D inference were the main obstacles. Connected-components post-processing was an effective, cheap way to suppress spurious predictions.

Methods

2D U-Net baseline with heavy augmentation (Albumentations)
3D patch-based U-Net for volumetric context
Class-imbalance handling via weighted/Dice + focal loss (placeholder)
Connected-components post-processing to remove spurious blobs

Results

2D vs 3D comparison documented with trade-offs (placeholder)
Post-processing reduced false positives (example result to update)

Lessons learned

Loss design matters more than architecture for extreme class imbalance
Volumetric context helps but is memory-bound; patching is a practical compromise

Limitations

Limited labeled data; results may not generalize across tissue types
3D inference is slow and memory-heavy

Next steps

Semi-supervised pretraining on unlabeled EM volumes
Active-learning loop to prioritize hard slices for labeling

Related projects

Featured 2024

Face Recognition Attendance System

Team-built attendance system using face recognition, with stranger rejection, reporting, and a dashboard.

Outcome: Example result to update — automated attendance capture with controlled false-positive rate.

Completed 2023

YOLO Object Detection & Tracking

Real-time object detection and multi-object tracking experiments comparing YOLO with several tracking algorithms.

Outcome: Example result to update — identified the best speed/accuracy trade-off for the target hardware.