Optical and X-ray fluorescence microscopy image pairs for field-of-view localization
This dataset holds the image pairs of XRF-to-Optical Field-of-View Localization with Vision Language Models: optical microscopy images and synchrotron X-ray fluorescence (XRF) maps acquired at the Advanced Photon Source (APS), the registrations that place each XRF map on its optical image, and the ground-truth boxes derived from them.
Collections
Controlled (data/controlled). One optical image of a tissue section, and five XRF maps of the same section measured at APS 2-ID-D (2idd_0023 to 2idd_0027). Each map was acquired at a region drawn on the optical image, which is tied to the scanning stage by four calibration points, so the drawn rectangles are the ground-truth boxes.
Challenging (data/challenging). Four specimen groups of archival formalin-fixed, paraffin-embedded lymph node and lung tissue from beagle dogs, obtained from the Northwestern University Radiobiology Archive (NURA) and prepared as described by Copeland-Hardin et al., Scientific Reports 13, 7806 (2023). For each group, one section was stained with hematoxylin and eosin and imaged with a Hamamatsu NanoZoomer slide scanner, and the adjacent section, mounted on an Ultralene membrane, was mapped by XRF microscopy at APS 2-ID-E (2xfm_*) and 8-BM-B (8bmb_*). Experts registered the 14 maps that lie within the four optical images from tissue and void patterns together with sample-trimming information. No fiducial markers were used. The nine maps that cover less than 20% of the optical image form the challenging dataset of the study; the five larger maps are reported separately.
Files
| File | Contents |
|---|---|
controlled.zip |
The controlled collection, unpacking to data/controlled/ |
challenging.zip |
The challenging collection, unpacking to data/challenging/ |
ground_truth.csv |
One row per map: its ground-truth box and the transform placing it |
Unpack both archives in one folder. Paths in the files start with data/.
data/controlled/
Controlled_6312-1.tif optical image
Controlled_6312-1.xml calibration points and the five drawn regions
img.dat/2idd_00NN.h5 XRF maps (5)
img.dat/2idd_00NN.xml map orientation
matches/visible_light_xrf_pairs.json the five optical-to-map registrations
data/challenging/
<specimen>_<date>_x2.5_z0.tif optical images (4)
img.dat/<map>.mda.h5 XRF maps (14)
matches/challenging_pairs.json each map's registration and ground-truth box
XRF maps are MAPS HDF5 files. The fitted elemental maps are arrays of channels by rows by columns, at MAPS/XRF_fits in the controlled maps and at MAPS/XRF_Analyzed/NNLS/Counts_Per_Sec in the challenging maps, with the channel names in MAPS/channel_names. The files also hold the beamline's scan records, including the stage positions of each scan. The study uses the phosphorus (P) channel.
Optical images are TIFF files. In Controlled_6312-1.xml, a region marker's box in optical pixels has its top-left corner at (RX + TopLeftX, RY + TopLeftY) and size RectWidth by RectHeight.
Ground truth
ground_truth.csv has one row per map, and pair_id identifies it in the matches files. box_x, box_y, box_w and box_h give the box's top-left corner and size in optical-image pixels, with the origin at the top-left corner of the optical image, x increasing to the right and y downward. fov_area_percent is the box area as a percentage of the optical image. vlm_paper_set gives each map's group in the study: controlled, challenging, or challenging with a field of view over 20% of the image.
To place a map as the study does, take its phosphorus channel, remove two pixels at each border, mirror it left to right if map_flip_h is true and top to bottom if map_flip_v is true, and then rotate it counter-clockwise by map_rotation_deg (as numpy.rot90 with k = map_rotation_deg / 90). For a controlled map, the oriented map then fills its box. For a challenging map, map_corners_optical_xy gives the optical-pixel positions of the oriented map's top-left, top-right, bottom-right and bottom-left corners. Three of them define the affine transform from map to optical pixels, including any mirroring in the registration, and the box is the axis-aligned bounding box of all four. optical_px_per_map_px is that transform's scale, and map_angle_deg the direction of the map's x axis in the optical image, in degrees from the optical x axis toward its y axis. For 2xfm_0009, 2xfm_0010, 2xfm_0012 and 2xfm_0013, the registration includes a rigid refinement fitted by matching each map's sulfur signal to the stained tissue.
Code
The code of the study is at https://github.com/AdvancedPhotonSource/visionlm-fov-localization.
Citation
Please cite XRF-to-Optical Field-of-View Localization with Vision Language Models when using these data.
- Downloads last month
- -