Revolutionizing Cancer Diagnosis: How Deep Learning Predicts Continuous Biomarkers with Unprecedented Accuracy

Digital pathology involves analyzing tissue specimens, often whole slide images (WSI), to predict genetic biomarkers for accurate tumor diagnosis. Deep learning models process WSI by breaking them into smaller regions or tiles and aggregating features to predict biomarkers. However, current methods primarily focus on categorical classification despite many continuous biomarkers. Regression analysis offers a more suitable approach, yet it must be explored. Some studies have used regression to predict gene expression levels or biomarker values from WSI but lack attention mechanisms or extensive validation. Further research is needed to compare regression and classification approaches in digital pathology to predict continuous biomarkers accurately.

Researchers from TUD Dresden University of Technology, University of Applied Sciences of Western Switzerland (HES-SO Valais), IBM Research Europe, Institute of Pathology, University Hospital RWTH Aachen, and many other institutes believe that regression-based deep learning (DL) surpasses classification-based DL. They introduce a self-supervised attention-based method for weakly supervised regression, predicting continuous biomarkers from 11,671 patient images across nine cancer types. Their approach significantly improves biomarker prediction accuracy and aligns better with clinically relevant regions than classification. In colorectal cancer patients, regression-based scores offer superior prognostic value. This open-source regression method presents a promising avenue for continuous biomarker analysis in computational pathology, enhancing diagnostic and prognostic capabilities.

The study uses regression-based deep-learning techniques to predict molecular biomarkers from pathology slides. The study excluded regression models from pathologist review due to unsatisfactory performance in quantitative metrics and the quality of generated heatmaps. The researchers investigated the prediction of lymphocytic infiltration from HE pathology slides in a large cohort of patients with colorectal cancer from the DACHS study. The image processing pipeline consisted of three main steps: image preprocessing, feature extraction, and classification-based attention attMIL for score aggregation, resulting in patient-level predictions. The study aimed to provide relevant prognostic information for colorectal cancer patients based on molecular biomarkers predicted from pathology slides.

The study utilizes regression-based deep-learning techniques to predict molecular biomarkers from pathology slides. The study employs the CAMIL regression method based on attention-based multiple-instance learning and self-supervised pretraining of the feature extractor. The research design includes using WSI for computational analysis of tissue specimen samples. The image processing pipeline consists of image preprocessing, feature extraction, and classification-based attention for score aggregation. The study focuses on predicting lymphocytic infiltration from HE pathology slides in a large cohort of patients with colorectal cancer.

The study developed a regression-based deep learning approach called CAMIL regression to predict Homologous Recombination Deficiency (HRD) directly from pathology images. They tested this approach across seven cancer types using The Cancer Genome Atlas (TCGA) cohorts and validated it externally using the Clinical Proteomic Tumor Analysis Consortium (CPTAC). CAMIL regression outperformed both classification-based DL and a previous regression method. It improved accuracy in predicting HRD status and showed greater class separability between HRD+ and HRD- patients compared to other approaches. Additionally, CAMIL regression demonstrated higher correlation coefficients with clinically derived ground-truth scores.

In conclusion, the study underscores the significant advancements offered by regression-based attMIL systems in digital pathology, particularly in predicting continuous biomarkers with clinical significance. Despite the limitations in the scope of the experiments and the inherent challenges in dealing with noisy labels and uncertainties in continuous biomarker measurements, the findings emphasize the potential of regression models in enhancing prognostic capabilities and refining predictions from histologic whole slide images. Further research should explore a broader spectrum of cancers and clinical targets while addressing the nuances between regression and classification approaches for more nuanced biological predictions. These insights pave the way for leveraging deep learning in precision medicine to its fullest extent.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and Google News. Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.