University of Missouri Research
Applied ML and computer vision research in histopathology and related domains, with emphasis on reproducibility and translation.
Research Projects
Whole-Slide Imaging Tools & Labeling
Built annotation and labeling tools for histopathology datasets using HistomicsTK and OpenCV to accelerate dataset creation across TB, Pap smear, and cell segmentation tasks.
- Reduced manual annotation effort by 100+ hours per dataset
- Integrated color deconvolution for stain separation
- Enabled scalable patch extraction and quality control
Color Deconvolution & Preprocessing for Stained Slides
Developed robust color deconvolution, normalization, and preprocessing pipelines to improve component separation and downstream ML accuracy on stained histology slides.
- Standardized stain variability across datasets
- Improved signal-to-noise for cellular structures
- Plug-and-play preprocessing modules for ML workflows
CNNs → Transformers Benchmarks
Benchmarked 20+ CNNs and Vision Transformers for Ziehl–Neelsen Mycobacteria detection, including ablations on augmentation strategies and ensemble methods.
- Top single model: RegNetX-8GF (95.99% accuracy)
- Best ensemble: DenseNet-169 + ResNet-34 (AUC=0.987)
- Found light/no color augmentation often outperformed heavy schemes
LLMs for Histopathology
Evaluated GPT vs Gemini for zero-/few-shot patch classification across AFB, CRIC, and BreaKHis datasets; explored prompt design and transfer learning.
- GPT-4.1 strongest; few-shot fine-tuning achieved 0.93 on AFB
- Multiclass prompting improved performance on CRIC
- AFB-tuned models transferred better to BreaKHis
Pap Smear Detection (ISBI PS3C)
End-to-end pipeline with self-supervised ViT pretraining (DINO) and advanced augmentation; competitive leaderboard performance.
- Placed 8th of 60 global teams (F1=0.7455)
- Strong results with DINO pretraining and tailored augments
- Iterative error analysis to refine hard-case performance
Neuroscience: fMRI Deep Learning (Algorithmic Mind Reader)
Decoded semantic thought representations from fMRI using neural networks and genetic algorithms; built interpretability and visualization pipelines.
- 20× better-than-random decoding accuracy
- Correlated brain regions with cognitive functions
- Visualization tooling for neuroscientific interpretability
Commercialization & Leadership (I-Corp & Coulter)
Completed two competitive commercialization programs, conducting 400+ stakeholder interviews and securing a $50K grant to advance translational AI in pathology.
- 400+ structured clinician and admin interviews
- $50K commercialization grant secured
- Investor-grade pitch materials and market validation
Publications
View all →LMMs for Histopathology
ICAIP 2025
Evaluation of commercial multimodal systems for patch-level classification across multiple datasets. Focus on prompt strategy and transfer behavior.
- GPT models consistently outperformed Gemini across datasets
- Few-shot fine-tuning with limited samples achieved strong performance
- Cross-domain transfer learning insights for medical imaging
Mycobacteria Detection Models
ICAIP 2025
Comparative study of CNNs and ViTs with augmentation and ensembles for Ziehl–Neelsen detection.
- Comprehensive benchmark of modern architectures
- Ensemble methods achieving state-of-the-art performance
- Systematic analysis of augmentation strategies