RADIX (Root Anatomy Deep-learning Image segmentation across species and platforms)

RADIX is a semantic segmentation model for plant root cross-section fluorescence microscopy. It uses a DINOv3 vision transformer encoder with a DPT decoder, and outputs per-pixel labels for 7 anatomical classes: Background, Epidermis, Aerenchyma, Endodermis, Vascular, Exodermis, and Cortex. Trained on Millet, Rice, Sorghum, and Tomato roots imaged on Olympus IX83, Cytation C10, and Zeiss LSM 970 microscopes.

---

Data Format

Each sample is a folder containing three grayscale 16-bit TIFs, one per fluorescence channel:

{sample}/
├── {sample}_DAPI.tif    # cell walls
├── {sample}_FITC.tif    # lignin
└── {sample}_TRITC.tif   # suberin

The filename prefix must match the folder name, and all three channels must share the same prefix. Both flat (data_dir/sample/*.tif) and two-level (data_dir/group/sample/*.tif) layouts are accepted; in the two-level case sample IDs become {group}/{sample}.

Images are normalized on the fly (1st-99.5th percentile per channel) and resized to 1024×1024 for inference; original resolution is preserved for downstream intensity measurements.

Optional ground-truth annotations use YOLO polygon format (class_id x1 y1 x2 y2 ..., normalized 0-1) with 6 raw classes: 0 Whole Root, 1 Aerenchyma, 2 Outer Endodermis, 3 Inner Endodermis, 4 Outer Exodermis, 5 Inner Exodermis. Rings are derived at load time by subtracting inner from outer polygons.

---

Running RADIX on New Data

The deployment entry point is predict.py. It runs inference with the trained DINOv3+DPT checkpoint and writes per-sample masks, visualizations, and a measurements CSV.

1. Open predict.py and set the two paths at the top:

   DATA_DIR  = Path("/path/to/samples/")   # folder of per-sample subfolders
   MODEL_DIR = Path("/path/to/run/")       # run dir (contains checkpoints/best-*.ckpt) or a .ckpt file

2. Run:

   python predict.py

3. Outputs land in {DATA_DIR}/predictions_dinov3_dpt/ (or a custom OUTPUT_DIR):

   predictions/{sample}.npy   # uint8 semantic argmax (0=bg, 1=epi, 2=aer, 3=endo, 4=vasc, 5=exo, 6=cortex)
   vis/{sample}.png           # input + Bio-7 overlay
   measurements.csv           # aerenchyma_ratio + per-region DAPI/FITC/TRITC means
   

Requires a CUDA GPU. Environment: pip install -r requirements.txt in a Python 3.11 conda env.

Evaluating against ground truth

If you have annotations and want IoU/Dice + paired GT-vs-prediction downstream comparisons, use run_eval_pipeline.py:

python run_eval_pipeline.py \
    --model-key timm_semantic \
    --checkpoint path/to/best.ckpt \
    --run-dir path/to/run/

This runs eval on the test and Zeiss oneshot splits, saves predictions as YOLO .txt files, then computes downstream measurements from both GT and predictions with correlation plots.

---

Polygon Editor

Interactive GUI for creating, reviewing, and correcting YOLO polygon annotations.

python polygon_editor.py --data-dir path/to/data/
python polygon_editor.py                             # no args: use Browse button

Modes (Mode dropdown)

Mode	Panels	Required folders	Purpose
Correct GT	3 (Original, GT, Prediction)	image/, annotation/, prediction/	Edit ground truth with predictions as reference
Correct Predictions	2 (Original, Editable)	image/, prediction/	Edit predictions, save into annotation/
Create GT	2 (Original, Editable)	image/	Draw annotations from scratch, save to annotation/

Class labels

Short names in the UI: Root, Aer, O.Endo, I.Endo, O.Exo, I.Exo.

Controls

Key	Action
N	Start drawing new polygon (click to add nodes)
B	Enter brush mode (default paint, Shift=erase, Ctrl+Scroll=size)
E	Edit selected polygon with brush (default erase, Shift=paint)
Enter / Space	Confirm drawing or edits
Escape	Cancel drawing or edits
Delete / Backspace	Delete selected vertex (edit mode) or polygon
R	Split a ring polygon into outer/inner
S	Save annotations
Ctrl+C	Copy selected reference polygon to editable panel
C	Copy all reference polygons to editable panel
1-6	Set class for new polygon
Ctrl+Z / Ctrl+Shift+Z	Undo / Redo
A / D / Left / Right	Previous / Next sample
H	Reset zoom, center all panels
Mouse wheel	Zoom
Middle/Right drag	Pan

Vertex editing

Drag vertices to move. Hover an edge to see a green "+" marker; click to insert a vertex. Select a vertex and press Delete to remove it.

Display sliders

Brightness (-100 to +100) and Gamma (0.1 to 3.0) affect the on-screen image only; the underlying raw TIFs are never modified.

Saving

Press S to save. Annotations are written to {data-dir}/annotation/ in YOLO polygon format, one file per sample: {Species}_{Microscope}_{Exp}_{Sample}.txt for the structured layout, or {sample}.txt for generic layouts. The folder is created if it does not exist.