Automated Detection of Microsatellite Status in Early Colon Cancer by Infrared Imaging Combined with Artificial Intelligence

The researchers of PRODI (Center for Protein Diagnostics, Ruhr-University Bochum, Germany) used tissue sections from the prospective multicentre AIO ColoPredictPlus 2.0 registry study to verify the novel label-free infrared (IR) imaging combined with artificial intelligence (AI) for tumour detection and classification of microsatellite status. The label-free IR images are measured by innovative Quantum Cascade Lasers (QCL) allowing fast measurements of several million spectra. They reported that AI integrated IR imaging can automatically classify unstained tumour sections precisely in less than 30 min. The approach is presented at proffered paper on gastrointestinal tumours, colorectal session during the ESMO Congress 2021 (16-21 September).

The authors explained that label-free QCL based IR imaging combined with deep learning provides spatially and molecularly resolved alterations of the genome and proteome in unstained cancer tissue sections. In particular, they classified IR images of tissue sections taken in 20 min with QCL IR microscopes by convolutional neural networks (CNN). An in-house developed segmenting CNN (U-Net) localises tumour regions and a second CNN (VGG-Net) subsequently classifies the microsatellite status. Endpoints were area under curve of receiver operating characteristic (AUROC) and area under precision recall curve (AUPRC).

The multicentre clinical cohort includes 491 patients, 100 tumour-free and 391 with tumour. Baseline characteristics of age, sex, stage, location, including BRAF mutation status were equally distributed among test cohorts.

The U-Net was verified on 491 patients from whom 294 serving as training dataset, 100 as test dataset and 97 as validation dataset. They result in an AUROC of 0.99 for the validation dataset. Tumour is thereby precisely spatially resolved in the sections.

The microsatellite status classification of the identified tumour regions was verified on 391 patients. 245 served as training dataset, 73 as test dataset and 73 as validation dataset. They reached currently an AUROC of 0.83 and an AUPRC of 0.64.

The authors expect further significant improvement during longer ongoing training phase. The approach is extended currently to prognosis and prediction of response which may contribute in the future as a new tool in precision oncology according to the study team.

Reference

385O – Großerüschkamp F, Schörner SM, Kraeft A-L, et al. Automated detection of microsatellite status in early colon cancer (CC) using artificial intelligence (AI) integrated infrared (IR) imaging on unstained samples from the AIO ColoPredictPlus 2.0 (CPP) registry study. ESMO Congress 2021 (16-21 September).