Todd Hollon

Introduces Prima, a visual foundation model for brain MRI trained at health-system scale for diagnosis, triage, and clinically grounded decision support.

FastGlioma is a computational pathology model for real-time detection of glioma infiltration at the surgical margin, outperforming the current standard of care.

Diffuse gliomas are classified using the molecular features. DeepGlioma predicts the molecular genetics of brain tumors within minutes of biopsy, in the operating room, to better inform surgical goals.

A deep learning workflow combining stimulated Raman histology with convolutional neural networks delivers near real-time intraoperative brain tumor diagnosis, matching pathologist accuracy while compressing turnaround from ~30 minutes to under 150 seconds.

Standard contrastive language-image pre-training can neglect objects in visual scenes. ItemizedCLIP forces models to learn and attend to all described items, resulting in better visual representations.

Recent visual agents can score well while using image tools unfaithfully-e.g., cropping irrelevant regions or ignoring tool outputs. CodeV represents tools as executable Python code and trains with Tool-Aware Policy Optimization (TAPO), using process-level rewards on visual tool inputs and outputs to improve both accuracy and faithful tool use on search and broader multimodal benchmarks.

Vision-language pre-training for volumetric MRI and CT is usually limited by radiologist-curated datasets. HLIP instead uses hierarchical attention over slice, scan, and study to pre-train on uncurated clinical data at scale, boosting performance on public brain MRI and head CT benchmarks.

HiDisc is a self-supervised learning method that leverages the inherent patient-slide-patch hierarchy of biomedical microscopy to learn stronger visual representations without explicit negative mining.

OpenSRH is the first public dataset of clinical stimulated Raman histology images from brain tumor patients, released alongside benchmarks to accelerate machine learning research for intraoperative brain tumor diagnosis.

NeuroVFM is a visual foundation model trained on 5.24M clinical MRI and CT volumes via health system learning, a paradigm that leverages uncurated data from routine care. Using a scalable volumetric joint-embedding predictive architecture, it delivers state-of-the-art radiologic diagnosis and report generation with interpretable visual grounding, surpassing frontier models in accuracy, triage, and expert preference while reducing hallucinations.

Using AI-quantified tumor infiltration from label-free optical microscopy of surgical margin tissue, combined with clinical, radiographic, and molecular features, a random forest model predicts sites of focal glioblastoma recurrence (validation AUC 80.3%). AI-derived infiltration was the strongest predictor, outperforming molecular features alone, pointing toward margin-guided precision adjuvant therapy for the highest-risk areas of disease.