IRM_InVitroMT: Microtubule dataset of interference microscopy images

This dataset contains 237 images of microtubule images acquired by interference microscopy and their corresponding masks. It support the work developed in the article by Sid'El Moctar et al (2026), where it was used to develop a Flow-Matching based deep learning architecture, CurvFlow, for segmenting microtubules. The dataset is split into 206 training images and 31 test images. Images were acquired using four microscopes (two Nikon Eclipse Ti-E and two Nikon Eclipse Ti-2) with 60× or 100× oil objectives (NA 1.49) Illumination was provided by either a CoolLED pE-300 or Nikon Intenslight E light source. Images were collected over several years and manually selected to ensure dataset diversity based on illumination patterns, signal-to-noise ratios, and microtubule coverage, size and orientation. Microtubules were manually annotated in ImageJ using the free handline tool (line thickness: 3 pixels for 60x objective, or 5 pixels for 100x objective). Annotations were converted to binary masks via a custom ImageJ macro, and ground-truth images were inverted to account for the bright field imaging modality.