Imaging and hyperspectral data from colorectal cancer tissue samples for multimodal machine learning models
- B. Borkovits
- E. Kontsek
- A. Pesti
- C. Antóny
- S. Gergely
- J. Slezsák
- A. Salgó
- B. Medgyes
- I. Csabai
- A. Kiss
- P. Pollner
2026-08-12
Colorectal cancer is one of the most common and deadly cancer types worldwide, with diagnosis and treatment outcomes heavily relying on histopathological assessment. While Whole Slide Images remain the gold standard for tissue data evaluation, techniques such as Fourier transform infrared spectroscopy may offer complementary molecular insights that are not visible through conventional staining. We present a multimodal dataset combining data recorded, using the two techniques, from 9 human colorectal cancer tissue microarrays containing 30–48 circular tissue cores extracted from 130 patients. The dataset includes Whole Slide Images of the whole tissue microarrays and individual mid-infrared spectroscopic measurements of the tissue cores, as well as RGB images of the measurement areas. The tissue cores consist of normal colon tissue, colorectal adenocarcinoma and colorectal liver metastasis tissue samples. Metadata files containing tissue core IDs and cancer labels were provided for convenient cross-modal matching as well as matching to patient IDs. Image quality and atmospheric effects influencing the infrared measurements were also evaluated and showcased in the article.