rajlab/condensatenet-v1

CondensateNet v1

Semantic segmentation model for detecting biomolecular condensates (puncta / stress granules) in fluorescence microscopy images.

Model Description

• Architecture: EfficientNetV2-S encoder with style-modulated Feature Pyramid Network (FPN) and dual output heads (binary mask + Cellpose-style flow fields)
• Parameters: 22.2M
• Input: Single-channel grayscale microscopy tile (320x320), percentile-normalized to [0, 1]
• Output: Binary segmentation mask logits + 2-channel flow field

Training

• Data: 245 validated image-mask pairs from 14 experiment types (fluorescence microscopy of various condensate-forming proteins)
• Split: 158 train / 26 val (stratified by experiment, with caps on overrepresented experiments)
• Loss: Focal + Tversky + flow MSE + differentiable calibration (soft-binned ECE)
• Optimizer: AdamW with linear warmup (5 epochs) + cosine annealing (120 epochs total, early stopped at epoch 94)
• Augmentation: Experiment-aware augmentation tiers (light/medium/heavy) with strong contrast scaling, CLAHE, intensity scaling (0.3x-2.0x), geometric transforms

Metrics (validation set)

Metric	Value
Dice	0.5164
F1	0.5164
IoU	0.3505

Inference metrics (full validation set, per-tile with threshold search)

Metric	Mean	Std
Dice	0.534	0.155
Precision	0.517	0.151
Recall	0.588	0.203

Usage

from condensatenet import CondensateNetPipeline

# Load from HuggingFace
pipeline = CondensateNetPipeline.from_pretrained("rajlab/condensatenet-v1")

# Segment an image
instances = pipeline.segment(image)
print(f"Found {instances.max()} condensates")

# Or load from a local download
pipeline = CondensateNetPipeline.from_local("/path/to/model")

Intended Use

Research tool for detecting biomolecular condensates in fluorescence microscopy. Not intended for clinical or diagnostic use.

Training Runs

This model was iteratively improved over 5 training runs:

Run	Changes	Val Dice
1	Baseline	0.503
2	+contrast augmentation	0.480
3	+calibration loss	0.509
4	+warmup/cosine/early stop (80 ep)	0.495
4b	Same, 120 epochs	0.516

Full training logs are available in the GitHub repository.