SEM Dendrite Segmentation

Author: Mohamad Dweik, Date: 25/02/2026

This project implements two different approaches for the semantic segmentation of dendrites in Scanning Electron Microscope (SEM) images: a Classic Computer Vision pipeline and a Deep Learning approach using YOLO.

Setup

Install the required dependencies using pip: pip install -r requirements.txt

1. Classic Computer Vision Approach

This deterministic pipeline processes images using classic techniques (Pre-processing -> Sauvola -> Morphology -> Skeletonization).

Open SEMDendriteSegmenter.py and scroll to the bottom of the file to choose how you want to run it:

Option A: Single Image Analysis (Debugging Mode) To see the step-by-step transformation of a single image:

1. Uncomment the lines loading a single image and calling segmenter.run_stages(img) and segmenter.show_stages(...).
2. Tip: Inside the run_stages() function, most of the intermediate steps are commented out by default to save memory. Just uncomment the specific stages["name"] = ... lines you want to visualize.
3. Run the script: python classic_cv.py

Option B: Batch Folder Analysis To process an entire folder and display a grid of the final results:

1. Comment out the single image lines at the bottom of the file.
2. Uncomment the segmenter.plot_all_overlays(...) line and make sure the folder_path points to your target images.
3. Run the script: python SEMDendriteSegmenter.py

2. Deep Learning Approach (YOLO)

Dataset Preparation

To improve segmentation of thin structures, the training images are split into overlapping tiles. Run this to generate the tiled dataset: python make_tiled_dataset.py

Training

To train the YOLO segmentation model (yolo26n-seg.pt): python train_yolo_dendrite_seg.py Results and the best model weights (best.pt) will be saved automatically in the runs/segment/ directory.

Inference on Large Images (Tiled Prediction)

For high-resolution images, we use a tiled inference approach with a voting mechanism to merge the masks. Important: Before running, open predict_tile.py and update the following variables with your local paths:

• weights: Path to your trained .pt model.
• source_folder: Path to the folder containing the images you want to predict on.
• out_dir: Path where you want the masks and overlays saved.

Once updated, run: python predict_tile.py

Inference on Standard Images

If your images are smaller and don't require tiling, use the standard prediction script. Like the tiled script, make sure to update your model and image paths inside the file first: python predict.py

Command Line Inference (Single Image)

If you prefer using the Ultralytics CLI directly on a single image: yolo segment predict model="runs/segment/yolo26_dendrite_tiled_final/weights/best.pt" source="path/to/new_image.png" imgsz=896 conf=0.15 iou=0.80 device=cpu save=True