Manthan Patel - Cell Segmentation & Tracking Showcase

Solo project by Manthan Patel. Update metrics and visuals as you progress.

Project Overview

Dataset(s): Fluo-N2DH-GOWT1 (track 01)
Role: Solo
Summary: Classical pixel-wise baselines (Naive Bayes, Logistic Regression, SVM) plus a deep learning pipeline featuring a ResNet34-based U-Net (Graduate Scope).

Segmentation Results

Model	Type	Mean IoU (Frame 0)	Notes
gnb	classical	0.872	Trained on Fluo-N2DH-GOWT1/01
logreg	classical	0.824	Trained on Fluo-N2DH-GOWT1/01
svm	classical	0.854	Trained on Fluo-N2DH-GOWT1/01
unet	deep	0.936	Trained on Fluo-N2DH-GOWT1/01
advanced_unet	deep	0.940	Trained on Fluo-N2DH-GOWT1/01

Preprocessing/features: sliding-window pixel classification with Sobel/LoG/texture channels; 50% FG sampling for class balance.
Validation Strategy: Models are trained on frames 0-2 and evaluated on held-out frames 3-4 to prevent data leakage and ensure unbiased reporting.
Repro: see scripts/run_final_benchmark.py for the exact pipeline used here.

Future Work / Closing the Gap

To further improve performance (aiming for <5% gap with top non-deep methods):

Ensemble Methods: Combine predictions from GNB, LogReg, and MLP using majority voting.
Advanced Features: Incorporate temporal features (optical flow) into the classical feature stack.
Post-processing: Apply watershed or CRF (Conditional Random Fields) to refine segmentation boundaries.
Data Augmentation: Augment training patches with rotation and flipping to improve robustness.

Tracking Results (Grad/Honors)

Tracker	Metric	Score	Notes
IoU linker	IDF1	0.737	Using U-Net segmentation
IoU linker	ID Switches	41	Over 92 frames
IoU linker	IDTP	1588	True Positive IDs
IoU linker	IDFP	664	False Positive IDs
IoU linker	IDFN	470	False Negative IDs

Heuristic IoU-based linking over per-frame masks.

Demo Videos / GIFs

Embed MP4/GIF assets committed under docs/assets/ or hosted elsewhere.

Tracking Demo

Reproduction Checklist

pip install -e .
python scripts/run_final_benchmark.py

Lessons Learned / Next Steps

Deep Learning Superiority: The Advanced U-Net (ResNet34) significantly outperforms classical baselines (~0.94 vs ~0.87 IoU).
Classical Baselines: GNB performed surprisingly well, beating Logistic Regression and SVM in this specific low-data regime.
Tracking: Simple IoU linking works well when segmentation is high-quality, but struggles with occlusions.

License

This project is released under CC BY-NC 4.0 (non-commercial, attribution required). Cite this repo for academic use.