SWITi Reproducibility

This repository reproduces the experiments for SWITi: Quantifying and Reducing Tiling Artifacts with Sliding Window Inner Tiling. SWITi is a test-time procedure for posterior image-splitting models that averages shifted tiled predictions to reduce seam artifacts without increasing the number of forward passes required for an MMSE estimate. The paper also introduces reference-free tiling-artifact metrics based on gradient-distribution tests.

The data and MicroSplit model weights are downloaded from Zenodo:

• Dataset: https://zenodo.org/records/22213583
• Model weights: https://zenodo.org/records/22214478.

SWITi has been implemented in CAREamics, an image restoration deep-learning library. (Currently in the feature branch mc/feat/switi-merge).

The gradient test and Fourier Ring Correlation (FRC) metrics are implemented in the repository juglab/TilArtMetrics.

Running the Reproduction

We recommend using uv to run the scripts from the repository root with uv run. See the uv documentation.

1. Download data and checkpoints.

   uv run scripts/download.py --output-dir switi

   This creates switi/data and switi/checkpoints.

2. Run baseline inner-tiling inference.

   uv run scripts/microsplit_inner_tiling_inference.py \
       --dataset HT_LIF24 \
       --data-root switi/data \
       --checkpoint-root switi/checkpoints \
       --prediction-root switi/results \
       --overlap 32 32 \
       --mmse-count 64

   This writes stitched predictions and inference_params.json under inner_tiling.

3. Run SWITi inference.

   uv run scripts/microsplit_switi_inference.py \
       --dataset HT_LIF24 \
       --data-root switi/data \
       --checkpoint-root switi/checkpoints \
       --prediction-root switi/results \
       --overlap 32 32 \
       --mmse-count 64

   For 3D datasets, pass a 3-axis overlap and --stride-z, for example --overlap 2 32 32 --stride-z 1.

4. Compute global image-quality metrics.

   uv run scripts/microsplit_compute_metrics.py \
       --dataset HT_LIF24 \
       --prediction-root switi/results \
       --prediction-subdir predictions_MMSE64 \
       --method inner_tiling \
       --data-root switi/data
   
   uv run scripts/microsplit_compute_metrics.py \
       --dataset HT_LIF24 \
       --prediction-root switi/results \
       --prediction-subdir predictions_MMSE64 \
       --method SWITi \
       --data-root switi/data

   This writes metrics.json and metrics_per_image.json beside each method's predictions.

5. Run FRC analysis.

   uv run scripts/run_frc_analysis_on_dataset.py \
       --dataset HT_LIF24 \
       --prediction-root switi/results \
       --prediction-subdir predictions_MMSE64 \
       --methods inner_tiling SWITi \
       --data-root switi/data \
       --ndim 2 \
       --output-dir switi/results

   This writes FRC reports, summary.csv, and per-channel FRC curve figures under frc.

6. Run the gradient permutation test.

   uv run scripts/run_gradient_test_on_dataset.py \
       --dataset HT_LIF24 \
       --prediction-root switi/results \
       --prediction-subdir predictions_MMSE64 \
       --methods inner_tiling SWITi \
       --data-root switi/data \
       --strip-width 1 \
       --output-dir switi/results

   This writes reports, summary.csv, and gradient_test_config.json under gradient_test.

7. Create example gradient-test overlay figures.

   uv run scripts/example_gradient_test_figure.py \
       --dataset HT_LIF24 \
       --prediction-root switi/results \
       --prediction-subdir predictions_MMSE64 \
       --data-root switi/data \
       --output-dir switi/results \
       --image-name <input-image-stem>

   The figure script reads the gradient-test reports and writes significance-overlay PNGs into the same gradient_test directory.

Expected Directory Structure

After the scripts above have run, the example directory should contain:

switi/
├── data/
│   └── HT_LIF24/
│       ├── inputs/
│       │   ├── train/*.tif
│       │   ├── val/*.tif
│       │   └── test/*.tif
│       ├── targets/
│       │   ├── train/*.tif
│       │   ├── val/*.tif
│       │   └── test/*.tif
│       └── stats.json
├── checkpoints/
│   └── HT_LIF24/
│       ├── BaselineVAECL_best.ckpt
│       └── config.yaml
└── results/
    └── HT_LIF24/
        └── predictions_MMSE64/
            ├── inner_tiling/
            │   ├── predictions.npz
            │   ├── inference_params.json
            │   ├── metrics.json
            │   └── metrics_per_image.json
            ├── sw_inner_tiling/
            │   ├── predictions.npz
            │   ├── inference_params.json
            │   ├── metrics.json
            │   └── metrics_per_image.json
            ├── gradient_test/
            │   ├── gradient_test_config.json
            │   ├── inner_tiling_gradient_report.json
            │   ├── SWITi_gradient_report.json
            │   ├── GT_gradient_report.json
            │   ├── summary.csv
            │   └── significance_overlay_<method>_channel<channel>.png
            └── frc/
                ├── inner_tiling_frc_report.json
                ├── SWITi_frc_report.json
                ├── summary.csv
                └── frc_curves_ch<channel>.pdf

• *.tif: input and target microscopy images for each split.
• stats.json: cached per-channel normalization statistics computed from the training data.
• BaselineVAECL_best.ckpt: pretrained MicroSplit checkpoint.
• config.yaml: training configuration used to reconstruct the model and prediction data configuration.
• predictions.npz: stitched predictions keyed by input image stem.
• inference_params.json: tile geometry, MMSE count, checkpoint path, and data path used for inference.
• metrics.json: dataset-average global image-quality metrics.
• metrics_per_image.json: global image-quality metrics for each predicted image.
• gradient_test_config.json: parameters used for the gradient permutation test.
• <method>_gradient_report.json: per-method gradient-test report.
• GT_gradient_report.json: gradient-test report for the ground truth reference, when included.
• significance_overlay_<method>_channel<channel>.png: example gradient-test overlay figures.
• <method>_frc_report.json: per-method FRC report.
• frc_curves_ch<channel>.pdf: FRC summary figure for one channel.

HPC Scripts

The hpc directory contains SLURM wrappers for the same workflow, which contain the arguments used to run the experiments:

hpc/microsplit_inner_tiling.sbatch
hpc/microsplit_switi.sbatch
hpc/microsplit_metrics.sbatch
hpc/run_gradient_test.sbatch
hpc/run_frc_analysis.sbatch