Wednesday, August 27, 2025

Artera, Owkin, on Digital Pathology in Cancer Management

Whole-slide imaging codes appear to be on hiatus at AMA CPT (here). but in the real world, the field is moving ahead steadily.  

Artera AI - Prostate

Artera AI won FDA de novo approval for AI prognostics based on prostate cancer slides.   See the press release here.   See Artera's diagnostics web page here.  See "Urology Times" here, which summarizes several published studies.

You can find the FDA webpage for DEN240068 here.  It can take a few months before FDA approval documents show up.  They will be interesting when they appear, as they'll reflect the FDA's current thinking on WSI & AI.

For now, Artera summarizes as, "the first and only AI-powered software authorized to prognosticate long-term outcomes for patients with non-metastatic prostate cancer."



Owkin - Breast Cancer (Barberis et al.)

Owkin publishes a paper on slide imaging for metastatic relapse risk in breast cancer.

See a Linked In article by Owkin's Victor Aubert here.  

Find the new publication, Barberis et al., in Nature Communications, here.  "Deep learning assessment of metastatic relapse risk from digitized breast cancer histological slides."


##

PubMed offered a "related title" to Barberis as being, "External validation of precisebreast, a digital prognostic test for predicting breast cancer recurrence, in an early-stage cohort from the Netherlands."  Westenend et al. 2025, here.  See more under "AI Corner."

##

AI CORNER

##

I asked Chat GPT to compare and contrast the two 2025 breast cancer WSI papers, Barberis 2025 and Westenend 2025.  

Below are direct AI output; check details before quoting.

##

Brief Compare

Both new AI-based pathology models address ER+/HER2− early breast cancer, but with different strengths for payor assessment. 

RlapsRisk-BC (Garberis et al.) predicts 5-year metastasis-free survival directly from routine slides, delivering clear incremental value over standard clinicopathologic factors (ΔC-index +0.05 overall, +0.08 in intermediate risk), which supports its potential role as a lower-cost triage or complement to genomic assays. 

PreciseBreast/PDxBR (Westenend et al.) predicts longer-term invasive disease-free survival and shows strong performance against MINDACT/MammaPrint clinical models, especially for grade 2 re-stratification, but adds less value beyond robust clinical-only models. 

For payors, the practical message is that RlapsRisk-BC demonstrates measurable utility versus clinical baselines and could reduce reliance on expensive molecular tests.  PDxBR’s appeal lies in refining ambiguous grade-2 cases and offering an alternative to genomic classifiers in longer-horizon risk planning.


###

Side-by-side for cancer experts: 

RlapsRisk-BC (Garberis et al., Nat Commun 2025) vs PreciseBreast/PDxBR (Westenend et al., Breast Cancer Res 2025)

Clinical question & endpoint.

  • RlapsRisk-BC targets ER+/HER2− early breast cancer and learns directly from H&E/HES whole-slide images (WSIs) to predict 5-year metastasis-free survival (MFS)—i.e., distant recurrence risk. Only ER+/HER2− patients were included, and analyses focus on 5-year MFS.

  • PDxBR also studies ER+/HER2− EBC, but its primary endpoint is invasive disease-free survival (IDFS) over longer follow-up (median 8.8 years); events include local/regional recurrence, distant metastasis, and death (drives higher event counts).

Takeaway: the endpoints differ (MFS vs IDFS), so absolute risk and discrimination cannot be compared one-to-one.

Populations & data

  • RlapsRisk-BC external validation (CANTO): n=1,169 ER+/HER2− patients; H&E and HES slides from multiple centers.

  • PDxBR (Netherlands NTH cohort): n=739 ER+/HER2−; treatments were typical of contemporary practice (endocrine 64%, chemo 29%, RT 68%) with 130 IDFS events (24 LRR/second primaries, 25 distant mets, 81 deaths).

Model architecture & inputs

  • RlapsRisk-BC is an end-to-end survival model trained to predict MFS directly from WSIs (no handcrafted features or molecular inputs). It later combines the WSI score with age, tumor size, grade, nodes, Ki-67 to form RR Combined. A calibration step (Uniform Piecewise Approximation) is required when deploying to a new lab—30 local slides (10 G1/15 G2/5 G3) align the score distribution across staining/scanner protocols.

  • PDxBR integrates an AI-derived grading signal (learned morphologic/protein-expression features from pathology images) with clinicopathologic variables (age, size, stage, nodal status). Imaging features contribute ~51% of the final score; gland/tubule architecture is the dominant image driver, while HR status and grade were dropped during feature selection (their signal captured by AI features).

Headline performance (external)

  • RlapsRisk-BC (CANTO, MFS): Adding the WSI score to clinical factors (RR Combined) improved Harrell’s C-index from 0.76→0.81 overall and 0.78→0.86 in the intermediate-risk subgroup (permutation p<0.05). It also outperformed PREDICT Breast (C-indices ~0.70–0.75).

    • With prespecified cutoffs (>5% predicted 5-yr MFS event = high risk), low- vs high-risk HRs were strong (e.g., HR 8.01 for RR Combined), with low-risk event rate ~1–2% at 5 years.

  • PDxBR (NTH, IDFS): C-index/AUC 0.71 overall; PDxBR Clinical 0.69, AI-grade 0.66, MINDACT clinical model 0.60. PDxBR significantly beat MINDACT (p<1e-5) and MammaPrint in the subset with genomic testing (n=252; p=0.0004). Notably, PDxBR vs PDxBR-Clinical was not statistically different (DeLong p=0.49).

Subgroups and re-stratification.

  • RlapsRisk-BC delivered the largest gains in grade 2 and intermediate-clinical-risk patients (sensitivity +21–28 points vs clinical-only), but premenopausal performance was limited.

  • PDxBR reclassifies histologic grade 2 (52% of cases) more evenly using the AI-grade and identifies high-risk cases missed by MINDACT/MammaPrint.

Cut-points & operating characteristics

  • RlapsRisk-BC used prespecified training-set cutoffs and validated with KM analyses; for CANTO, high risk was >5% 5-yr MFS event probability. At this setting, low-risk had ~1.3–1.6% 5-yr metastasis events.

  • PDxBR uses a fixed score threshold of 58, yielding HR 3.05, sensitivity 0.70, specificity 0.59, NPV 0.90, PPV 0.27 for IDFS in NTH (vs prior validation HR 4.4; 6-year IDFS).

Interpretability

  • RlapsRisk-BC: Pathologist-reviewed tile clusters showed increased risk signals for high tumor cell content, nuclear pleomorphism, mitoses, massive/trabecular architecture, and decreased risk with fibrosis or vascular structures, supporting biologic face validity.

  • PDxBR: Quantified feature attributions highlight gland differentiation/tubule formation as the largest single contributor to the score, with nuclear pleomorphism and mitotic figures also important; imaging features explained ~51% of the score.

Implementation & generalizability

  • RlapsRisk-BC explicitly addresses domain shift with the 30-slide calibration (UPA) and shows concordant performance across H&E and HES; authors note weaker performance in premenopausal women and call for head-to-head vs genomic assays and prospective trials.

  • PDxBR deployed on a different national cohort with longer follow-up and demonstrated stable discrimination, significant superiority to MINDACT, and practical grade-2 re-stratification; however, combining images+clinical did not significantly outperform clinical-only in this dataset.

How these results might be used

  • If your clinic struggles with grade-2 and “intermediate-risk” calls: Both tools help, but RlapsRisk-BC showed added value over a strong clinical model (ΔC-index +0.05 overall; +0.08 in intermediate risk) and better sensitivity/specificity tradeoffs at a 5-yr MFS cutpoint.

  • If you need a model aligned to broader IDFS events and longer horizons: PDxBR provides solid 8.8-year IDFS discrimination (C-index 0.71) and beats the MINDACT clinical model, with clear, interpretable morphologic drivers and a fixed pragmatic threshold (58).

  • Genomics positioning: RlapsRisk-BC explicitly proposes WSI as a pre-screen or complement to genomic assays, while PDxBR shows cases misclassified by MammaPrint/MINDACT that its image+clinical approach captures.

Key contrasts in one glance

  • Endpoint focus: MFS (RlapsRisk-BC) vs IDFS (PDxBR).

  • Strength vs clinical baseline: RlapsRisk-BC adds significant incremental value to a multi-variable clinicopath model; PDxBR’s combined model did not significantly exceed its clinical-only model in NTH.

  • Deployment needs: RlapsRisk-BC requires site calibration (30 slides); PDxBR uses a fixed score threshold (58) with standard WSI inputs.

  • Interpretability: Both lean on classic histopathology (glands/tubules, pleomorphism, mitoses); RlapsRisk-BC also flags stroma/vascular correlates of lower risk.

Bottom line for practice & trials

  • For adjuvant-therapy decision support in ER+/HER2− EBC, RlapsRisk-BC currently shows clear incremental prognostic value over clinical factors for 5-yr distant relapse, especially in grade-2/intermediate-risk cases—useful where the question is chemotherapy vs no chemo in the near term. Prospective utility studies (and operationalizing the calibration step) are the next barrier.

  • PDxBR offers a long-horizon, event-inclusive measure that outperforms MINDACT-style clinical modeling and reclassifies ambiguous grades, which may matter for extended endocrine therapy or surveillance planning; however, its combined model didn’t beat clinical-only here, so payer/HTA arguments may hinge on grade-2 re-stratification and MINDACT/MammaPrint deltas rather than pure incremental AUC vs clinical.

 ##