scSplit: Bringing Severity Cognizance to Image Decomposition in Fluorescence Microscopy

Overview

This is the official implementation of scSplit: Bringing Severity Cognizance to Image Decomposition in Fluorescence Microscopy, published at NeurIPS 2025. Paper link.

Installation

1. Install mamba.
2. Execute installation.sh.

After installation, one needs to do mamba activate split_hpc to activate the environment and one can subsequently start to train or evaluate the models.

Training

For $Gen_i$ networks

python /home/ashesh.ashesh/code/DiffSplitting/split.py -c /home/ashesh.ashesh/code/DiffSplitting/config/hagen_indiSplit.json -enable_wandb

For $Reg$ network

python /home/ashesh.ashesh/code/DiffSplitting/time_prediction_training.py  --config=/home/ashesh.ashesh/code/DiffSplitting/config/ht_t24_time_predictor.json -enable_wandb

Evaluation

Evaluation is done by running the notebooks. For some of the tasks, it can take more time and so we execute the notebooks as scripts.

Synthetically Summed Inputs

In this case the argument --mixing_t_ood=0.5 fixes the desired mixing-ratio in the inputs.

python notebooks/evaluate_notebook.py  --mmse_count=10  --notebook=/home/ashesh.ashesh/code/DiffSplitting/notebooks/EvaluateJointIndi.ipynb --ckpt=2502/BioSR-joint_indi-l1/5 --mixing_t_ood=0.5  --ckpt_time_predictor=2502/BioSR-UnetClassifier-l2/4 --outputdir=/group/jug/ashesh/indiSplit/notebook_results/ --training_rootdir=PATH_TO_WHERE_ALL_MODELS_ARE_SAVED

Directly Imaged Inputs

python notebooks/evaluate_notebook.py --enable_real_input=true --ckpt=2502/HT_LIF24-joint_indi-l1/60 --mmse_count=10  --notebook=/home/ashesh.ashesh/code/DiffSplitting/notebooks/EvaluateJointIndiRealInput.ipynb --ckpt_time_predictor=2502/HT_LIF24-UnetClassifier-l2/3 --training_rootdir=PATH_TO_WHERE_ALL_MODELS_ARE_SAVED

Datasets

The publicly available datasets were utilized in our study. We adopted the same train/validation/test splits as those employed by previous methods focused on semantic unmixing. For convenience and clarity, we also provide these train/validation/test splits here.

• Hagen et al.
• BioSR
• HTT24
• HTLIF24
• PaviaATN

Pre-trained models

• Hagen et al.
• BioSR
• HTT24
• HTLIF24
• PaviaATN

For evaluating other scSplit variants.

In addition to evaluating scSplit, the same checkpoints can be used to evaluate other scSplit variants mentioned in the paper. One needs to make following changes to the arguments while evaluating the notebooks:

• For $\text{scSplit}_{0.5}$, provide --infer_time=False --use_aggregated_inferred_time=False --use_hardcoded_time_for_inference=0.5 and donot provide ckpt_time_predictor.
• For $\text{scSplit}_{-agg}$, provide --use_aggregated_inferred_time=False. `