Todd Hollon, Joseph R Novello M.D. and Alfredo Quinones-Hinojosa M.D. Research Professor of Neurosurgery, Michigan Medical School
Watch RecordingTodd Hollon’s AI journey began not with a dataset or an algorithm, but with a view through a surgical microscope that had not changed in a century. When a neurosurgeon opens a patient’s skull to remove a brain tumor, the tissue inside looks normal. The tumor does not form a discrete ball — it infiltrates the surrounding brain, making the boundary between cancer and healthy tissue invisible to the naked eye. “This all looks very normal,” Hollon told the audience. “And that’s how it is for the majority of brain cancers.” A surgeon must decide, in real time and under enormous pressure, whether to keep cutting or stop — a decision with irreversible consequences for a patient who, unlike someone with cancer in a limb, cannot simply have the affected region removed.
That clinical impasse became Hollon’s research question: could computer vision make a decision that human eyes cannot? His lab had already developed a specialized optical microscope — a stimulated Raman histology imager — capable of producing high-resolution images of brain tissue in real time during surgery. But as Hollon was quick to recognize, an image is not a decision. What the lab needed was an AI system that could translate an image into a clinically actionable prediction, and doing that responsibly — before AI had “really infiltrated into healthcare,” as he put it — required building from clinical principles rather than borrowing methods from adjacent fields.
The work began in 2016, before TensorFlow had fully matured and before AI had become a staple of medical research. The team built what Hollon describes as a two-stage architecture: a large encoder model, analogous in structure to GPT, trained to understand the visual language of brain tissue, followed by a downstream “slide score” model trained specifically to estimate what proportion of a given specimen is infiltrated with tumor. The choice to use a large pretrained encoder rather than a task-specific model from scratch reflected a deliberate judgment: general visual understanding first, clinical specificity second. The output was a continuous estimate — not a binary call, but a gradient — giving surgeons a graded signal they could act on rather than a yes or no.
Validation was built into the project architecture from the start. Rather than testing on a held-out slice of the training data, Hollon’s team released FastGlioma to collaborators at UCSF, NYU, and Vienna and required that it perform consistently across institutions, tumor types, and patient populations — including both adults and children — without retraining at each site. The results were framed not in model accuracy metrics but in clinical outcomes: the likelihood of patients receiving optimal surgical resection increased, and the rate of surgical errors went down. That framing — outcomes over benchmarks — was a deliberate choice about how AI results should be reported when the audience is a surgeon weighing a real decision, not a researcher evaluating a model.
Hollon was candid about where the work goes next. The team is accumulating data to answer the harder forecasting question: at the moment of surgery, how much overall clinical improvement has the operation bought the patient? “That’s a predictive task,” he said, “a forecasting task which is generally harder.” He is also developing non-invasive imaging approaches — MRIs and CTs — to better understand prognostic factors before a patient ever reaches the operating room. His message to students was pointed: the next generation of AI development will not come from more language modeling or more GPUs. It will come from new ideas, applied in interdisciplinary settings, to problems that the large technology companies are not focused on. “We need brilliant young minds at universities to work on these problems.”