Featured Publications and Research
Researchers: Onur C. Koyun, PhD, Yongxin Guo, Ziyu Su, PhD, Hao Lu, PhD, and Metin N. Gurcan, PhD
The Problem
- Oncotype DX (ODX) is clinically valuable for prognosis and therapy decisions in early-stage HR+ breast cancer, but it is expensive and can have long turnaround times, limiting access
- There is a need for an accessible alternative using routine pathology (H&E WSIs)
The Research
- The study proposes AnchorMIL, a framework to predict ODX from H&E whole-slide images
- Approach: an anchored regression–classification mechanism that produces both a continuous risk estimate and a binary risk classification
- Validation data: TCGA-BRCA (public multi-center) and OSU (institutional) cohorts.
- Performance: AUC 0.89 (TCGA-BRCA) and 0.86 (OSU)
- Interpretability: The framework highlights features associated with aggressive disease biology (e.g., lymphovascular invasion and comedonecrosis)
The Impact
- Offers a scalable, lower-cost risk stratification approach that could reduce reliance on genomic assays, accelerate treatment decisions, and potentially improve equity in care
Interested in the full study? Read More
Researchers: Yongxin Guo, Ziyu Su, PhD, Onur C. Koyun, PhD, Hao Lu, PhD, Metin N. Gurcan, PhD
The Problem
- Oncotype DX (21-gene recurrence score) guides therapy decisions in HR+/HER2−
breast cancer, but is expensive and has access barriers - Need a cost-effective alternative using routinely available pathology (H&E WSIs).
- WSIs are gigapixel-sized → computationally heavy; weak labels make learning
difficult
The Research
- Proposed BPMambaMIL: a bio-inspired prototype-guided weakly supervised MIL framework integrating Mamba mechanism + prototypical guidance to predict ODX score intervals from H&E WSIs
- Clinical (in-house) dataset: 398 HR+/HER2− WSIs (267 low-risk, 131 high-risk), using ODX=25 as threshold
- Performance (in-house): AUC = 0.839, +5.61% over baseline MambaMIL.
- High-risk classification: accuracy 0.714 vs 0.419 (baseline) in clinical dataset
- Generalizability: validated via cross-dataset evaluations on two external public datasets (incl. TCGA-BRCA and Camelyon16) vs six SOTA baselines
The Impact
- Improves ODX risk assessment from routine pathology → potential lower-cost alternative to genomic assays
- Prototype-guided design aims to capture complex histopathologic patterns and improve robustness across scenarios
- Stronger identification of high-risk group (clinically most actionable bucket) compared to baseline
Interested in the full study? Read More
Researchers: Abbas Alili, PhD, Ava M. McKane, Muhammet F. Demir, MD, Cynthia L. Emory, et al.
The Problem
- For patients with suspected metastatic bone disease (MBD), delays in referral from radiology reports to orthopedics can negatively affect outcomes; manual report review is a major bottleneck
- Delayed specialist involvement can prevent timely interventions before complications such as fractures occur
The Research
- SHIELD is an NLP framework designed to automatically triage radiology reports for orthopedic referral need.
- Model: Fine-tunes RadBERT-RoBERTa-4m to classify reports into three categories: No Referral / Referral / Referral–High Risk.
- Data & period: Two academic centers, Jan 2014–May 2025; MBD cohort includes 245 eligible patients plus a 245-patient control cohort.
- Labeled set: 555 expert-labeled reports; 15% hold-out test, 85% train/val with 5-fold stratified CV.
- Explainability: Uses Llama-3.1-8B-Instruct to generate natural-language rationales for predictions.
- Pipeline components:
- Binary (referral vs non-referral): Accuracy 100%, AUC 1.00
- 3-class: Overall accuracy 89.52%, F1 99.20% for “No Referral”
- “Fail-safe”: does not classify high-risk cases as “No Referral.”
The Impact
- Retrospective timeline analysis suggests it could reduce an average 109.6-day referral delay to a computational triage step executed in ~1–3 minutes
- The explainability layer may improve clinician trust and adoption
Interested in the full study? Read More
Researchers: Hao Lu, PhD, Muhammet F. Demir, MD, Gabriella I. Puchall, Zian Shang, Shalaka Chavan, Aaron C. Moberly
The Problem
- Diagnosing middle-ear conditions (e.g., acute otitis media, AOM) from otoscopy is often subjective, increasing the risk of diagnostic errors
- Although deep learning models using otoscopy videos are promising, many pipelines rely on manual selection of the “best frame” (Most Informative Frame, MIF), which limits real-world scalability
The Research
- LUCID proposes a systematic method to automatically identify the Most Informative Frame (MIF) from otoscopy videos.
- Using 713 videos, the study identifies three main drivers of frame informativeness:
- 1. Eardrum visibility, 2) Eardrum coverage, 3) Image clarity
- Pipeline components:
- ResNet-50 (trained on >38,000 labeled frames) to score eardrum visibility
- BC-AdvCAM (weakly supervised) for eardrum segmentation and coverage estimation
- A blur/focus metric tailored to otoscopy image quality
- These signals are fused into an informative score to rank frames.
- Key result: Automatically selected frames achieve diagnostic performance comparable to expert-chosen frames; using the top-4 frames (instead of a single frame) further improves diagnostic accuracy
The Impact
- Automates the major bottleneck—manual MIF selection—making video-based otoscopy AI more scalable for clinical workflows
- Supports “expert-level” frame selection and suggests that multi-frame (top-4) usage can improve robustness and accuracy
Interested in the full study? Read More
Researchers: Ziyu Su, PhD, Yongxin Guo, Khalid Niazi, PhD, and Metin Gurcan, PhD
The Problem
- The Oncotype DX (ODX) test is expensive and hard to access
- Underserved communities often don’t get it
- There’s a need for cheaper, more accessible options
The Research
- Developed Deep-BCR-Auto – an AI tool that predicts ODX risk from regular H&E pathology slides
- Tested on TCGA-BRCA and OSU datasets (n=1006 and n=465)
- Reached AUROC scores of 0.827 and 0.832
- Performed well across different ages, races, and cancer types
- Most mistakes happened in borderline cases (ODX scores between 20-30)
The Impact
- A low-cost alternative to gene tests
- Doesn’t require pathology input
- Helps make personalized treatment more available to everyone
- Show strong accuracy and potential for real-world clinical use
Interested in the full study? Read More
Researchers: Usman Afzaal, Ziyu Su, PhD, Khalid Niazi, PhD, and Metin Gurcan, PhD
The Problem
- Accurate cancer grading is key for treatment decisions
- Manual grading by pathologists can be subjective and inconsistent
- There’s a need for faster, more objective tools
The Research
- Created an AI model that grades clear cell renal cell carcinoma (ccRCC) from H&E slides
- Trained and tested on two datasets (TCGA and WFBCCC)
- Achieved high accuracy: AUROC of 0.904 on TCGA, 0.854 on WFBCCC
- Performed consistently across different races and institutions
The Impact
- Offers a fast, accurate, and scalable way to grade ccRCC
- Reduces variability in manual assessments
- Supports pathologists and improves patient care
- Moves AI one step closer to routine use in cancer diagnosis
Interested in the full study? Read More
Researchers: Mostafa Rezapour, PhD, Metin Gurcan, PhD, Aarthi Narayanan, PhD, and Wyatt Mowery
The Problem
- Rapid, accurate infection detection is critical in Ebola outbreaks
- Tools like RNA-Seq and NanoString are used, but their reliability and consistency needed comparison
The Research
- Machine learning applied to gene expression data from Ebola-infected primates
- Strong agreement found between RNA-Seq and NanoString platforms
- OAS1 identified as a consistent biomarker distinguishing infected vs. Uninfected samples
- RNA-Seq detected additional antiviral genes, offering a broader immune profile
The Impact
- Validates both RNA-Seq and NanoString for Ebola diagnostics
- Positions OAS1 as a promising diagnostic marker
- Lays groundwork for faster, more accurate tests and deeper insight into immune response
Interested in the full study? Read More