nuclear_speckles_analysis

This repository trains a UNet-style image-to-image translation model on cropped nuclei from nuclear_speckles_data.

• Input: cropped DAPI (CH0) nucleus image
• Target: cropped Gold (CH2) nucleus image
• Task: predict Gold crops from DAPI crops

Data Pipeline

Training uses cached, filtered single-cell crops generated from:

• nuclear_speckles_data/IC_corrected_images
• nuclear_speckles_data/Preprocessed_data/single_cell_profiles/*annotated*.parquet

datasets/dataset_00/utils/CropCacheBuilder.py builds the crop cache and manifest when needed. Plate, well, site, and channel metadata are parsed from IC_corrected_images filenames split by _:

• field 1: plate
• field 2: well
• field 3: site
• field 4: channel (CH0 for DAPI input, CH2 for Gold target)

Normalization

Before samples are passed through the model, both input and target crops are normalized by their channel max intensity (dtype max, for example 255 for uint8 and 65535 for uint16) in datasets/dataset_00/utils/ImagePreProcessor.py.

Training

Main entrypoint:

mlflow run . -e train_model

Fast smoke run:

uv run train.py --epochs 1 --n-trials 1 --max-train-batches 2 --max-eval-batches 2 --enable-image-savers 0

Loss and Metrics

• Optimization loss: L1
• Logged metrics: L1, L2, PSNR, SSIM

Metrics are implemented in metrics/ and logged through callbacks/Callbacks.py.

Outputs

• Crop-level prediction artifacts can be logged each epoch via callbacks/utils/SaveEpochCrops.py.
• Whole-FOV reconstruction and patch stitching are intentionally removed.