VP – Cell Segmentation & Tracking Showcase

Project Overview

This project implements a complete segmentation pipeline on the Cell Tracking Challenge dataset using both classical machine-learning models and a deep-learning U-Net.

Dataset

• Fluo-N2DH-GOWT1 (2D fluorescent microscopy)
• Training & evaluation using Track 01 ST (silver-truth) masks

Goals

• Implement the full classical baseline (Naive Bayes, Logistic Regression, SVMs, MLP)
• Add a graduate-level deep learning segmentation model (U-Net)
• Evaluate using Mean IoU (Jaccard) via MySEGMeasure.py wrapper
• Produce a clear, fully reproducible research pipeline and results page

Workflow Summary

1. Data Preparation

• Downloaded the Fluo-N2DH-GOWT1 dataset using:

python scripts/setup_data.py Fluo-N2DH-GOWT1 --splits training test

• Data is stored under:

artifacts/datasets/Fluo-N2DH-GOWT1/
  training/
    01, 01_ST, 01_GT, ...
    02, 02_ST, 02_GT, ...
  test/

• ST masks are used for training and evaluation.

2. Classical Segmentation Pipeline

• Extract per-pixel features from a 5×5 sliding window around each pixel.
• Construct a balanced training set (foreground/background sampling).
• Train the following classical models on these handcrafted features:
  • Naive Bayes
  • Logistic Regression
  • Linear SVM
  • RBF SVM
  • MLP (1–hidden-layer neural network)

3. Deep Model (U-Net)

• Implemented a small U-Net in PyTorch.
• Trained the U-Net directly on raw image patches + ST masks.
• Evaluated predictions using the same IoU / SEGMeasure evaluation tools, for a fair comparison with classical models.

Segmentation Results (Track 01, ST Masks)

All results below are reported on Fluo-N2DH-GOWT1, Track 01, ST masks, using Mean IoU / Jaccard.

IoU Bar Chart

A visual comparison of Mean IoU across all models:

This bar chart summarizes the segmentation performance of all classical baselines and the deep U-Net on the same track and silver-truth masks.

Qualitative Results (Example Frame t000)

Below is a qualitative comparison for time frame t000.

Raw Image (t000)

Ground Truth (ST Mask)

Predicted Mask

Classical Pipeline Details

Feature Extraction

• For each pixel, a 5×5 neighborhood is extracted from the raw image.
• The 5×5 patch is flattened into a 25-dimensional feature vector.
• Optional normalization / scaling can be applied.
• Training data is formed by sampling pixels from:
  • Foreground regions (cells) in ST masks
  • Background regions

Classical Models Overview

Each model produces a probability or score per pixel, which is converted into a binary segmentation mask and then evaluated via IoU and SEGMeasure.

Deep Model: U-Net

Architecture (Simplified)

Input (1 × H × W)
        ↓
Encoder: [Conv → ReLU → Downsample] × 4
        ↓
      Bottleneck
        ↓
Decoder: [Upsample → Skip Connection → Conv → ReLU] × 4
        ↓
Output (1 × H × W logits)
        ↓
Sigmoid → Binary mask

Skip connections from encoder to decoder help preserve fine-grained spatial detail, which is crucial for accurate cell boundaries.

Training Settings

Even under these modest settings, the U-Net produces sharp and realistic masks, especially for medium-to-large cells.

Reproduction Checklist

This section describes how to reproduce the main experiments from a fresh clone of the repo.

1. Environment Setup

conda env create -f environment.yml
conda activate cse488-cell-tracking
pip install -e .

2. Download Dataset

python scripts/setup_data.py Fluo-N2DH-GOWT1 --splits training test

3. Train Classical Models (Example Commands)

# Naive Bayes
python scripts/train_naive_bayes.py Fluo-N2DH-GOWT1 --track 01

# Logistic Regression
python scripts/train_logreg.py Fluo-N2DH-GOWT1 --track 01

# SVM with RBF kernel
python scripts/train_svm.py Fluo-N2DH-GOWT1 --track 01 --kernel rbf

# Linear SVM
python scripts/train_svm_linear.py Fluo-N2DH-GOWT1 --track 01

# MLP (shallow neural net)
python scripts/train_mlp.py Fluo-N2DH-GOWT1 --track 01

Each script saves the trained model into artifacts/models/ (e.g., nb_track01.pkl, logreg_track01.pkl, etc.).

4. Train U-Net

python scripts/train_unet.py Fluo-N2DH-GOWT1 --track 01 --epochs 10 --batch-size 4 --lr 0.001 --model-path artifacts/models/unet_track01.pt

5. Evaluate Models

Example evaluation commands (with verbose per-label IoU):

# Naive Bayes
python scripts/eval_seg.py Fluo-N2DH-GOWT1 --track 01 --model-type naive_bayes --model-path artifacts/models/nb_track01.pkl --verbose

# Logistic Regression
python scripts/eval_seg.py Fluo-N2DH-GOWT1 --track 01 --model-type logreg --model-path artifacts/models/logreg_track01.pkl --verbose

# Linear SVM
python scripts/eval_seg.py Fluo-N2DH-GOWT1 --track 01 --model-type svm --model-path artifacts/models/svm_linear_track01.pkl --verbose

# MLP
python scripts/eval_seg.py Fluo-N2DH-GOWT1 --track 01 --model-type mlp --model-path artifacts/models/mlp_track01.pkl --verbose

# U-Net
python scripts/eval_seg.py Fluo-N2DH-GOWT1 --track 01 --model-type unet --model-path artifacts/models/unet_track01.pt --verbose

The script reports Mean IoU (Jaccard) and per-label IoU values, along with a summary Jaccard index consistent with MySEGMeasure.py.

Conclusion

This project successfully reproduces the required classical segmentation baseline and extends it with a U-Net deep-learning model on the Fluo-N2DH-GOWT1 dataset.

• The MLP achieved the highest Mean IoU (≈0.866) among all classical models.
• The U-Net achieved ≈0.811 Mean IoU, performing extremely well on large cells but missing some very small objects due to limited training and lack of augmentation.

Overall, the repository provides a fully reproducible, modular, and extensible segmentation system, ready for future improvements such as:

• Stronger data augmentation
• Dice / Focal / Tversky losses
• Multi-track training
• Temporal tracking and linking of cell identities over time