# Regression-based Deep-Learning predicts molecular biomarkers from pathology slides

**Authors:** Omar S. M. El Nahhas, Chiara M. L. Loeffler, Zunamys I. Carrero, Marko van Treeck, Fiona R. Kolbinger, Katherine J. Hewitt, Hannah S. Muti, Mara Graziani, Qinghe Zeng, Julien Calderaro, Nadina Ortiz-Brüchle, Tanwei Yuan, Michael Hoffmeister, Hermann Brenner, Alexander Brobeil, Jorge S. Reis-Filho, Jakob Nikolas Kather  
**arXiv:** [2304.05153](https://arxiv.org/abs/2304.05153)  
**Date:** 2023-04-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.05153v1/

## Key points

- Deep learning (DL) has shown promise in predicting biomarkers from cancer histopathology
- Most existing approaches focus on categorical labels rather than continuous measurements
- Researchers aimed to investigate whether regression-based DL outperforms classification-based DL in predicting continuous biomarkers directly from images
- Developed a new self-supervised attention-based weakly supervised regression method
- Evaluated using a large dataset of 11,671 patients across nine different cancer types
- Tested method on multiple clinically and biologically relevant biomarkers, including the homologous repair deficiency (HRD) score
- Regression significantly improved accuracy of biomarker prediction compared to classification
- Regression-based scores provided higher prognostic value than classification-based scores in colorectal cancer patients
- Regression-based DL enhances accuracy and interpretability of biomarker prediction
- Anonymized patient samples were used with approval from ethical boards
- Image data consisted of raw whole-slide images scanned by an Aperio ScanSlide scanner and preprocessed for analysis
- Two types of clinical targets were analyzed: continuous variables with known cutoffs and continuous variables without known cutoffs requiring categorization by splitting at the median
- Model pipeline involved image preprocessing, feature extraction, and two approaches for score aggregation: classification based attention attMIL and regression based attMIL.
- Study presents a promising alternative for analyzing continuous biomarkers in computational pathology using an open source regression approach.

## Summary

Deep learning (DL) has shown promise in predicting biomarkers from cancer histopathology, but most existing approaches focus on categorical labels rather than continuous measurements. In this study, the researchers aimed to investigate whether regression-based DL outperforms classification-based DL in predicting continuous biomarkers directly from images. To achieve this, they developed a new self-supervised attention-based weakly supervised regression method and evaluated it using a large dataset of 11,671 patients across nine different cancer types. The researchers tested their method on multiple clinically and biologically relevant biomarkers, including the homologous repair deficiency (HRD) score, which is a widely used pan-cancer biomarker. The results showed that using regression significantly improved the accuracy of biomarker prediction compared to classification. Additionally, regression-based scores provided higher prognostic value than classification-based scores in a large cohort of colorectal cancer patients. This suggests that regression-based DL not only enhances the accuracy of biomarker prediction but also improves its interpretability. The study also included an ethics statement confirming that anonymized patient samples from several academic institutions were used with approval from ethical boards. The image data and cohorts consisted of raw whole-slide images (WSIs) scanned by an Aperio ScanSlide scanner and preprocessed for analysis. Two types of clinical targets were analyzed: continuous variables with known clinically relevant cutoffs for categorization and continuous variables without known cutoffs requiring categorization by splitting at the median. The research team described their model pipeline which involved image preprocessing, feature extraction and two approaches for score aggregation: classification based attention attMIL and regression based attMIL. The entire pipeline allowed for patient level predictions based on WSI data. Overall, this study presents a promising alternative for analyzing continuous biomarkers in computational pathology using an open source regression approach. By improving both accuracy and interpretability, regression based DL has the potential to advance biomarker analysis and contribute to personalized cancer care.