PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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:2304.05153v1cs.CVcs.AI

TL;DR

Most deep-learning approaches predict categorical labels even when biomarkers are continuous, and earlier regression approaches lacked systematic comparison with classification. The paper evaluates CAMIL regression across cancer types and biomarkers, reporting improved prediction, interpretability, and survival stratification over classification in the evaluated settings.

  • Problem

    Most published approaches classify biomarkers, although many have continuous ground-truth values, and prior regression methods lacked systematic comparison with classification.

    Source: Introduction · pp. 2–3

  • Method

    The authors developed CAMIL regression, combining self-supervised feature learning with attention-based multiple instance learning to predict continuous biomarkers from pathology images.

    Source: Introduction · p. 3

  • Results

    CAMIL regression improved biomarker prediction accuracy and separability over classification, with better interpretability and survival prediction in the evaluated settings.

    Source: Discussion · p. 7

  • Takeaways & Limitations

    Regression-based prediction is a promising alternative for continuous biomarker analysis in computational pathology.

    Source: Discussion · p. 7

    Experiments covered a selected set of tumors and clinical targets, some without external test sets containing the same clinical information; model hyperparameters were not optimized.

    Source: Discussion · p. 7

Abstract

from arXiv · show

Deep Learning (DL) can predict biomarkers from cancer histopathology. Several clinically approved applications use this technology. Most approaches, however, predict categorical labels, whereas biomarkers are often continuous measurements. We hypothesized that regression-based DL outperforms classification-based DL. Therefore, we developed and evaluated a new self-supervised attention-based weakly supervised regression method that predicts continuous biomarkers directly from images in 11,671 patients across nine cancer types. We tested our method for multiple clinically and biologically relevant biomarkers: homologous repair deficiency (HRD) score, a clinically used pan-cancer biomarker, as well as markers of key biological processes in the tumor microenvironment. Using regression significantly enhances the accuracy of biomarker prediction, while also improving the interpretability of the results over classification. In a large cohort of colorectal cancer patients, regression-based prediction scores provide a higher prognostic value than classification-based scores. Our open-source regression approach offers a promising alternative for continuous biomarker analysis in computational pathology.

Introduction

WSI-based deep learning can predict molecular biomarkers, but most methods cast continuous biomarker measurements as categories. This study compares classification and regression across cancer types and introduces CAMIL regression.

Introduction

  • WSI-based deep learning predicts gene alterations, microsatellite instability, and gene-expression patterns, and several biomarker-prediction applications have regulatory approval.Pathology images also enable correlations between genetic changes and morphology.Source: Introduction · p. 2
  • Weakly supervised prediction pairs gigapixel WSIs with one patient-level sequencing ground truth, while attention-based multiple instance learning aggregates tile features for patient-level predictions.Tiles can include connective tissue or fat that may not contribute to biomarker predictability.Source: Introduction · p. 2
  • Most published approaches classify categorical labels, although many biomarkers have continuous values that are binarized before modeling.Prior studies have also used custom category thresholds or percentile cutoffs for evaluation.Source: Introduction · pp. 2–3
  • Binarization loses information, and prior continuous-value regression had not been systematically compared with classification and required more validation.The authors identify this as a gap in evidence for regression on pathological images.Source: Introduction · p. 3
  • The study introduces CAMIL regression, combining self-supervised learning with attention-based multiple instance learning, and compares it with classification across cancer types and biomarkers.The comparison examines biomarker predictability, interpretability, and prognostic capability.Source: Introduction · p. 3

Results

CAMIL regression predicted HRD and biological-process biomarkers from histology, and outperformed classification in several benchmark, interpretability, and survival analyses. In colorectal cancer, its continuous scores improved survival risk stratification over classification.

Regression predicts HRD from histology

  • CAMIL regression achieved HRD-status AUROCs above 0.70 in 6 of 7 TCGA cancer types and was evaluated in external CPTAC cohorts.CPTAC AUROCs ranged from 0.62 in LUSC to 0.96 in UCEC.Source: Regression predicts HRD from histology · p. 3

Regression outperforms the state-of-the-art classification-based approach

  • CAMIL regression outperformed Graziani regression and CAMIL classification for HRD AUROC in all 7 TCGA cancer types, with statistically significant differences in 5.It also had higher R2 scores than Graziani regression in all 7 TCGA and 4 CPTAC cohorts; the ablation attributed much of Graziani’s inferior performance to its optimizer.Source: Regression outperforms the… · p. 4

Regression predicts key biological process biomarkers from histology

  • CAMIL regression averaged a 4% AUROC increase over classification across five biological-process biomarkers, outperforming it in 8 of 34 TCGA instances and tying in the rest.The biomarkers span tumor-cell proliferation, stromal fraction, and immune-cell measures.Source: Regression predicts key biological… · p. 5

Regression improves interpretability of biomarker predictions from histology

  • In blinded review, regression heatmaps were most interpretable in 34 of 42 cases (81%), with sharper delineation of lymphocyte-rich regions than classification heatmaps.Classification heatmaps were preferred in 6 cases, and the approaches were judged similar in 2.Source: Regression improves interpretability… · pp. 5–6

Regression-based biomarkers improve survival prediction in colorectal cancer

  • In DACHS colorectal cancer patients, regression scores significantly stratified survival for 3 of 5 biomarkers univariately and 2 of 5 multivariately, with stronger hazard ratios than classification.The analyses used long-term follow-up data and overall survival as the endpoint.Source: Regression-based biomarkers improve… · p. 6

Discussion

The authors argue that continuous biomarker prediction is often reduced to classification, and report that CAMIL regression improves prediction, interpretability, and survival stratification in their tested settings. They frame the study as proof of principle while noting limits in tumor and target coverage, external testing, and hyperparameter optimization.

Discussion

  • The study reports direct evidence that CAMIL regression outperforms classification for predicting continuous biomarkers, which are often categorized by traditional deep-learning methods.The authors describe the approach as expanding pathology-based biomarker prediction to a key clinical application area.Source: Discussion · p. 7
  • CAMIL regression improved biomarker prediction and separability over CAMIL classification, particularly for HRD, and showed better external generalization than Graziani et al.’s regression approach.The authors attribute the latter difference predominantly to Graziani et al.’s optimizer causing convergence to the mean.Source: Discussion · p. 7
  • Regression heatmaps were preferred for interpretability in 81% of tumor-infiltrating-lymphocyte cases and improved survival risk stratification for immunologic biomarkers versus classification.The authors state that these biomarkers were selected for their prognostic capabilities.Source: Discussion · p. 7
  • The experiments covered a selected set of tumors and clinical targets; not all targets had an external test set with the same clinical information, and model hyperparameters were not optimized.The authors characterize the work as proof of principle and call for broader cancer and target analyses.Source: Discussion · p. 7

Materials and Methods

The study analyzes 11,671 whole-slide images across continuous clinical and biological targets, comparing classification and regression models with site-aware evaluation. Its pipeline combines image preprocessing, self-supervised tile features, attention-based aggregation, statistical and survival analyses, and heatmap review.

Ethics statement

  • Ethics approvals were obtained through DACHS and the Technical University Dresden; retrospective analyses of anonymized TCGA and CPTAC samples did not require formal ethics approval.The overall analysis was approved under BO-EK-444102022.Source: Ethics statement · p. 7

Image Data and Cohorts

  • 11,671 whole-slide images were analyzed, including HRD with a clinically relevant cutoff and five biological-process biomarkers categorized at their medians when no relevant cutoff was known.The biological targets were assessed across seven cancer types; TCGA-LIHC lacked TIL regional-fraction data.Source: Image Data and Cohorts · p. 8

Model description

  • The pipeline preprocesses slides, extracts tile features, then aggregates them with classification- or regression-based attention multiple instance learning to produce patient-level predictions.RetCCL converts each preprocessed tile into a 2048-dimensional feature vector using a self-supervised, contrastive-clustering model fine-tuned on 32,000 WSIs.Source: Model description · p. 8

Experimental setup and implementation details

  • Models used patient-level, site-aware five-fold cross-validation, Balanced MSE for imbalance, and customized classification or regression heads; comparisons used matched patients and included an adapted Graziani regression baseline.The site-aware splits grouped patients by originating hospital to limit bias from site-specific histological features.Source: Experimental setup and implementation… · p. 9

Statistics and endpoints

  • AUROC compared classification and regression using the same binarized labels, while HRD score distributions and paired tests assessed additional performance differences.The study also used independent univariate and multivariate Cox analyses on external DACHS data to assess survival prediction with age, sex, and tumor stage as covariates.Source: Statistics and endpoints · p. 10

Visualization and explainability

  • Attention heatmaps were generated from trained models for TCGA-BRCA LISS predictions and blindly reviewed by a pathologist for interpretability.The review compared heatmaps for 42 patients, producing 84 maps in total.Source: Visualization and explainability · p. 11

Data and Code availability

  • The authors report that source code for preprocessing, classification, regression, and attention heatmaps is available under an open-source license.The passage also identifies public sources for TCGA and CPTAC slides and molecular data.Source: Data and Code availability · p. 11

Figures

The figures depict the workflow, HRD model comparisons, biological-process biomarker comparisons and attention maps, and external survival analyses. Their captions specify the cohorts, model outputs, metrics, and visual encodings.

Figure 1

  • Figure 1 traces image preprocessing and tile-level feature extraction into attMIL modeling heads, with performance metrics and study cohorts also summarized.The workflow uses a ResNet50/RetCCL feature extractor and separate classification and regression heads.Source: Figure 1 · p. 12, Figures · p. 13

Figure 2

  • Figure 2 compares AUROC distributions for three HRD prediction approaches on internal TCGA and external CPTAC test sets, alongside normalized score distributions for UCEC.Transparent violin instances indicate non-significant AUROCs; statistical tests are shown on the plots.Source: Figure 2 · p. 14, Figures · p. 15

Figure 3

  • Figure 3 maps AUROC deltas between CAMIL regression and classification across five biological-process biomarkers and cancer types; positive deltas favor regression.A separate panel shows classification and regression attention heatmaps for a TCGA-BRCA LISS slide.Source: Figure 3 · p. 15

Figure 4

  • Figure 4 presents univariate and multivariate Cox analyses of model predictions on external DACHS colorectal cancer samples, with hazard ratios and confidence intervals.The multivariate analyses include tumor stage, age, and sex as covariates; stars denote p < 0.05.Source: Figure 4 · p. 16

Funding

The funding passages list research grants and institutional support, alongside a disclosure of one author’s external financial and advisory roles.

Funding

  • The listed support includes German federal, European Union, NIHR, and other grant programs for JNK and institutional or fellowship support for additional authors.The passage names funding schemes including DEEP LIVER, PEARL, CAMINO, SWAG, ODELIA, and GENIAL.Source: Funding · p. 21
  • J.S.R.-F. reports consultancy fees, advisory-board memberships, and a board directorship outside the scope of the study.The named organizations include Goldman Sachs, Bain Capital, REPARE Therapeutics, and Paige.AI.Source: Funding · p. 21

Competing Interests

JNK reports consulting relationships and lecture honoraria; J.S.R.-F. reports partial funding from three organizations.

  • JNK reports consulting services for Owkin, Panakeia, and DoMore Diagnostics, and lecture honoraria from MSD, Eisai, and Fresenius.Source: Competing Interests · p. 21
  • J.S.R.-F. reports partial funding from the Breast Cancer Research Foundation, a Susan G. Komen Leadership grant, and an NIH/NCI grant.Source: Competing Interests · p. 21

Supplements

The supplements compare regression and classification for HRD and biological-process biomarkers, report regression-model comparisons, and describe cohort availability for validation and survival analyses.

Raw results and statistics

  • HRD comparisons use AUROC/AUPRC and ANOVA across site-aware TCGA training cohorts and CPTAC validation cohorts; regression-versus-classification tests use dependent one-sided t-tests.Source: Raw results and statistics · pp. 22–24
  • Positive percentages in Suppl. Table 4 indicate larger median-peak separation for regression; negative percentages indicate larger separation for classification.Source: Raw results and statistics · p. 25
  • AUROC alone can misrank regression models: CAMIL regression had a wider prediction range and higher R², while AUROC indicated better performance for Graziani regression.On CPTAC-LUAD, prediction ranges were (10, 55) for CAMIL regression and (29, 34) for Graziani regression.Source: CAMIL regression has better… · p. 27
  • R² 0.29 (1.39E-40) for CAMIL regression versus 0.16 (3,06E-22) for Graziani regression on external CPTAC-LUAD, although AUROC favored Graziani.Source: CAMIL regression has better… · p. 27
  • The CAMIL regression ablation used TCGA-BRCA site-aware splits and varied dropout, optimizer, epoch count, and kernel-based balancing.TCGA-BRCA was selected because both regression models had statistically significant AUROCs across all five folds with similar ranges and low variance.Source: CAMIL regression has better… · p. 28
  • Suppl. Figure 2 plots predicted score distributions against continuous ground-truth scores and reports R² and Spearman’s ρ across five site-aware test splits.Source: CAMIL regression has better… · pp. 34–35

Data availability

  • The TCGA-CRC cohort had 625 available whole-slide images from 632 patients with clinical information; DACHS had 2,297 patients with overlapping images and survival data.DACHS provided the external colorectal cohort for overall-survival prediction.Source: Data availability · p. 39
  • HRD analyses used CPTAC external-validation cohorts, while biological-process biomarker analyses used an external colorectal cohort from DACHS.The listed CPTAC cohorts were LUAD, LUSC, PAAD, and UCEC.Source: Data availability · pp. 39–40

Related papers

Loading 2304.05153v1…