PCUK2023 FTIR Prostate TMA Dataset

Dataset Contents This dataset contains 1045 Fourier Transform Infrared (FTIR) images of prostate tissue micro-array (TMA) cores from a set of 21 TMA slides and a total of 181 unique patients. Each TMA core in this dataset has associated with it the following data: FTIR Image in hdf5 format [256 x 256 x 1479] Greyscale 'chemical' image calculated as the area under the Amide I spectral band [256 x 256 x 3] Binary tissue/background mask [256 x 256 x 3] H&E image from a serial section, aligned with the FTIR [1024 x 1024 x 3] Core-level cancer/benign labels Core-level Gleason grading Patient ID Patient-level TNM cancer staging The file 'master_sheet.xlsx' contains a reference of every TMA core's name and its associated metadata. Files in this dataset are in the format s<slide_id>_c<core_id>.<file_extension>. The file 'wavenumbers.h5' contains the exact wavenumber values corresponding to each of the 1479 channels in the FTIR images. Tissue Annotation A subset of 227 TMA cores spanning 100 patients in this dataset have sparse pixel-level annotations of the following tissue types: Tissue Type Annotation Colour R / G / B Normal epithelium lime green [0, 255, 0] Normal stroma purple [128, 0, 128] Cancerous epithelium magenta [255, 0, 255] Cancer-associated stroma blue [0, 0, 255] Corpora amylacea orange [255, 165, 0] Blood red [255, 0, 0] Crushed tissue cyan [0, 255, 255] Immune infiltration yellow [255, 255, 0] Necrotic debris with immune infiltration dark red [127, 0, 0] Tissue classification results for this annotated subset from a variety of machine learning models including Patch-CNNs, U-Net, ViT, and RandomForest are reported in the following work: "Spatial-Spectral Deep Learning for Prostate Cancer Tissue Classification in Infrared Spectroscopy" https://doi.org/10.1021/acs.analchem.5c04765 Tissue Origins Prostate tissue in this dataset was accessed via the Manchester Cancer Research Cancer Biobank (10_NOCL_02) with ethical approval granted by the South Manchester Research Ethics Committee (ref: 22/NW/0237). Spectral Acquisition IR images were captured using an Agilent Agilent Cary 620 FTIR Microscope at 15× objective. FTIR images are padded to a uniform 256px height and width with a channel dimension of 1479 spanning wavenumbers 949->3800cm-1. Spectra in this dataset are entirely unprocessed. Further information on spectral acquisition is available in the following works: "Spatial-Spectral Deep Learning for Prostate Cancer Tissue Classification in Infrared Spectroscopy" https://doi.org/10.1021/acs.analchem.5c04765 "Identification of at-risk prostate cancer patients using Fourier transform infrared spectroscopy and machine learning" https://doi.org/10.1117/12.3048498 Usage Spectral data are saved in hdf5 file format, and so can be accessed using any code library that can iterface with this. An example in Python using the h5py package might look like: import h5py f = h5py.File('pcuk2023/spectral/s02_c004.h5','r') spectra = f['spectra'][:,:,:] mask = f['mask'][:,:] f.close() print(spectra.shape) >>> (256, 256, 1479) Contact Lyra O'Leary | [e-mail] | data curation Peter Gardner | [e-mail] | project lead