Center-for-Living-Systems/mt_structure_classification: v0_newton_2026

Image Classification of Microtubule-GUV Structures This repository contains a complete workflow for performing image classification on grayscale microscopy images of microtubule-GUV (Giant Unilamellar Vesicle) structures. The project includes model training, evaluation, visualization, and inference on new test datasets. What This Project Does The pipeline takes paired fluorescence microscopy TIFFs (GUV channel + microtubule channel) and: Indexes image pairs — matches GUV and MT files by naming conventions Preprocesses — 2D max-projection, pad/crop to 512x512, background subtraction (median filter), percentile normalization Segments GUVs — detects individual vesicles using Cellpose (deep learning instance segmentation) and/or Hough circle detection, with filtering by eccentricity, area, and MT signal Crops MT patches — extracts 96x96 pixel patches from each detected GUV Classifies — fine-tuned EfficientNet-B0 (also supports ResNet18, ConvNeXt-Tiny) classifies patches into 5 MT structure classes using focal loss A trained model is provided in the models/ folder. This is the checkpoint used for the results presented in the paper (trained on the paper's data). Use it for inference with run_prediction.py by passing --model-path models/best_model.pth (and --label-map models/label_map.json if needed). Pipeline overview (paper results) | Step | Description | CLI / Package | |------|-------------|----------------| | 1 | Index GUV/MT image pairs | scripts/run_preprocessing.py (with step 2) | | 2 | Background + channel statistics | scripts/run_preprocessing.py | | 3 | Segment GUVs (Hough and/or Cellpose) | scripts/run_segmentation.py (with step 4) | | 4 | Crop 96×96 MT patches per object | scripts/run_segmentation.py | | 5 | Train classifier (EfficientNet, focal loss) | scripts/run_training.py | | 6 | Predict on folder of crops | scripts/run_prediction.py |