USCAP 2025

Tempus is advancing precision medicine through the practical application of artificial intelligence in healthcare. We are pleased to share our latest scientific and clinical research findings during the United States and Canadian Academy of Pathology 2025 Annual Meeting.

Schedule a meeting with us

Abstracts
Online-only

Abstracts

March 25, 2025

Time
9:30am-12:00pm EDT

Abstract #
840

Poster #
197

Presenter
Jacob Gordan

Laboratory Development Test Validation of Deep Learning Histogenomic Models to Predict MSI Status in Multiple Cancer Types

We developed a model to predict total nucleic acid (TNA) yield from H&E slide images to support next-generation sequencing (NGS) workflows, addressing the issue of insufficient TNA yield leading to NGS testing failures. The model was trained using data from NGS samples collected from January-June 2023, incorporating factors like cell count, sample age, tissue site, and procedure type. It was validated on a temporal test set (July-Sept 2023) and evaluated post-deployment (July-Aug 2024) across various cancers and procedures. The model achieved high AUC values (0.83 for DNA QNS and 0.89 for TNA > 1000 ng) and positive predictive values (PPV) in both validation and deployment datasets. Performance was consistent across different cancer types and procedures, with AUC values above 0.80 for DNA QNS and above 0.84 for TNA > 1000 ng. The model helps optimize lab workflows by identifying cases likely to fail due to low TNA yield early, allowing for additional sample collection, and by flagging cases with excess TNA yield for more efficient use of slides.

Time
9:30am-12:00pm EDT

Abstract #
331

Poster #
192

Presenter
Kshitij Ingale

Validation and Deployment of H&E Image Based Model Predicting Total Nucleic Acid Yield in Multiple Cancer Types

We developed an AI model to predict microsatellite instability-high (MSI-H) status from (H&E) whole-slide images (WSIs) for prostate, colorectal, and endometrial cancers, validated following CAP/CLIA standards. The dataset included a large number of WSIs for model development and validation sets enriched for MSI-H cases. Attention-based multiple instance learning models were trained and evaluated for analytical accuracy, precision, sensitivity, and specificity. We showed that the models had high predictive accuracy (AUC) for MSI status and maintained over 80% concordance in MSI-H predictions between original and rescanned slides. The models met all predefined acceptance criteria, and the prostate cancer model is now deployed internally to assist pathologists in triaging patients for further confirmatory testing. Future deployments of these algorithms are planned.

Time
1:00pm-4:30pm EDT

Abstract #
333

Poster #
239

Presenter
Bo Osinski

Modular Validation of Lymphocyte Detection and Tumor and Stroma Segmentation Models to Accurately Predict Tumor Infiltrating Lymphocytes from H&E Images in Metastatic Non-Small Cell Lung Cancer

We developed an AI model to accurately quantify tumor-infiltrating lymphocytes (TIL) in non-small cell lung cancer (NSCLC) to predict patient response to immune checkpoint inhibitor (ICI) therapy. Our AI model consists of two sub-modules: one for lymphocyte detection and another for tumor and stromal region segmentation, validated using high-quality ground truth (GT) from IHC-derived labels and pathologist consensus. The lymphocyte model was trained on 280 slides, with IHC staining for T/B lymphocytes, and the tumor/stroma model was trained on annotations from 260 H&E slides. Results show strong correlation between predicted lymphocyte counts and IHC-derived GT counts (CCC 0.91), with high F1 scores for cell detection and classification across various tissue sites. The models accurately measure TIL density, which could guide personalized treatment plans for NSCLC patients by informing ICI therapy decisions.

Online-only

Presenter
Bo Osinski

Comparison of IHC-to-H&E Registration vs H&E-Only Ground Truth Methodologies for Evaluating Lymphocyte Detection AI Models on H&E images of Non-Small Cell Lung Cancer

Developing AI models for tumor-infiltrating lymphocyte (TIL) detection is challenging because establishing lymphocyte ground truth (GT) is subject to human variability in the manual labeling of H&E stained tissues. H&E slides from mNSCLC were annotated for two datasets: one on H&E only and one using IHC-derived labels, and lymphocytes were annotated by three pathologists. Inter-pathologist agreement for lymphocytes was much higher when annotating on IHC than H&E. Comparing AI model performance on IHC-derived and H&E-only datasets, we found the model overpredicts cells relative to H&E-only GT but is better correlated with IHC-derived GT. The model has improved cell and lymphocyte detection on IHC-derived GT.

Note that this was accepted, but we are not presenting it due to time constraints.

Schedule a meeting with us

We'll be in touch shortly.