Virtual Biopsies in One Second: How a Seven-Year-Old CycleGAN Is Beating Newer Models in the Clinic

Host Nathan Labenz frames the second paper as a stark before-and-after. Today a biopsied tissue sample is fixed, sliced into very thin sections, plated on a slide, dyed with chemical reagents and finally examined under a microscope, a destructive process that takes about eight hours. In the new approach there is no slicing at all: a 3D imaging technology captures the whole tissue, a slice is pulled out of the volume, and a model virtually stains it.

Abraham confirms the summary and stresses why the delay matters. An eight-hour process is simply impractical for applications like surgery, where a surgeon removing a brain tumor wants to know immediately whether any tumor remains. Faster alternatives exist, at around 30 minutes, but he says they produce low-quality images that are hard to interpret.

Labenz adds that the output could also feed a classifier, and that a speedup of roughly two orders of magnitude would probably mean far more of this imaging gets done. Both men see the stained image as a bridge to automated screening as well as to human review.

Abraham explains that qOBM, quantitative oblique back illumination microscopy, was developed at a lab at Georgia Tech that collaborated on the paper. It is slide-free, meaning it needs no sectioning, and also label-free, meaning it needs no dyes. He points out that the most damaging steps are not the stain itself but the formalin fixation and paraffin embedding needed to cut thin slices.

Some slide-free methods still rely on stains, and those cannot be used on living tissue, since nobody wants to dye functioning tissue inside the body. Because qOBM needs no labels, it can image in vivo, and the Georgia Tech lab has already miniaturized it into a handheld probe. Combined with AI, he says, this becomes a powerful diagnostic system for surgical settings.

Labenz connects this to a familiar experience: relatives told only that the doctor thinks they got the whole tumor. The hope, Abraham agrees, is to put a probe into the surgical site, capture a 3D image, virtually slice and stain it, and check for remaining tumor in line with the operation.

The dataset was tiny by modern standards: roughly 2,358 qOBM tiles and 1,737 H&E tiles, each 512 by 512 pixels. These came from only a few rats and a few human patients, so the tiles are sub-regions of a small number of tissue samples. Abraham agrees with Labenz that this is really small data.

He explains why collecting more is so hard in medicine. Getting rat data means growing tumors and sacrificing many animals, and getting human data means waiting for enough brain tumor surgeries and hoping excess tissue can be used for research. Standard H&E slides are plentiful because clinics already store them, but a novel imaging technology has no such archive, and bringing a new technology into the clinic brings its own challenges.

The deeper problem is that ground truth cannot be matched pixel for pixel. The fresh tissue imaged by qOBM must then be fixed, embedded, sliced and stained, and that processing changes the tissue, so the two images are slightly different. That makes this an unpaired image-to-image translation task, which is where CycleGAN comes in.

Abraham says CycleGAN, introduced around 2016-2017, is the most common approach to unpaired translation. In medicine, he notes, 2017 still counts as new, whereas in AI a six-month-old method can feel ancient. He tried newer alternatives, including a contrastive-learning approach and diffusion models, and none worked as well.

CycleGAN, by contrast, worked very well right out of the box and with limited data. He is not completely sure why, but says that in his experience these models do not seem to overfit and generalize decently as long as the input data is high quality. His hypothesis is that newer methods carry implicit assumptions that do not match the properties of histology images.

He admits he was surprised, having expected to be working with diffusion models by now. His lesson is that the newest technique is not always the best for a given task, and that revisiting older methods can pay off.

Abraham walks through GANs: a generator tries to produce realistic images and a discriminator tries to tell real from generated, each improving against the other. He calls GANs the first primitive form of generative AI, used for things like StyleGAN faces up to about 2020. CycleGAN adds a second generator and discriminator, so one generator turns a qOBM image into a fake H&E image and the other turns that back into a qOBM image.

The trick is comparing the reconstructed qOBM image to the original. For the round trip to succeed, the content has to be preserved in the intermediate H&E image, which is why the method is called cycle-consistent. Abraham is candid that this is not a guarantee, and that you are essentially hoping content is maintained. At inference, only the first generator is needed, even though four models are trained.

He shares a key insight from the work. In H&E the cell nuclei are dark, but in qOBM, especially tumor nuclei, they appear bright, which can lead the model to mark the wrong regions as nuclei while still satisfying cycle consistency. The team's fix was simply to invert the qOBM images so nuclei become dark, making the easiest solution also the correct one.

Abraham describes this as understanding the psychology of the model, and says much of the project was data processing and cleaning. He calls himself a big believer in data-centric AI, arguing that even basic models work well on good datasets.

Labenz notes that wherever the system was confused, it flagged tissue as a potential tumor rather than missing real tumors. Abraham says this was not explicitly designed, and he traces the errors to structures that look different in intact qOBM tissue than in processed H&E: continuous capillaries, red blood cells that fall out during processing, and white matter bundles.

Ideally there would be neither false positives nor false negatives. He points out that false positives matter in brain surgery too, since they could lead to removing healthy, functioning tissue. More data and better approaches may help close that gap.

Training is cheap: each model trains on a single A100 in a few hours, in part because the datasets are small. Getting to the clinic is the expensive part. The team had only one healthy human image, which the model handled well despite being trained mostly on tumors, and it used only grade two and three astrocytomas, so other tumor types, more specimens and eventually clinical trials are needed.

Abraham adds that the handheld probe is still at the research stage and will need its own validation. He stresses that medical testing is intensive for good reason, since the technology would be used in life-and-death situations. He returns to his main takeaway from his PhD: the most advanced models do not help without high-quality data, or as his advisor liked to say, garbage in, garbage out.

Abraham describes his ties to the open research community EleutherAI and to Stability AI. He says the former inspired MedARC's open, collaborative style, while he is a part-time employee at Stability, which supports his work with compute and funded a fellowship for his PhD. He says nothing exclusive is returned to Stability and that the company supports many academic projects.

He argues that medical AI is a two-way street. The original U-Net was developed for cell segmentation before spreading widely, including into diffusion models, and CLIP, he says, began as a Stanford model on image-text radiology reports before OpenAI scaled it up. That is part of why organizations like Stability are interested in supporting medical AI research.

On radiology, he sees AI as a tool for now, useful as a starting point despite hallucinations, with doctors still in the loop. He cites papers suggesting a combined AI-doctor system can outperform either alone, and mentions his team's recent work fine-tuning Stable Diffusion on chest X-rays to help train doctors. He is also excited about AI finding molecular biomarkers in H&E images that people cannot see, while saying that unless reliability improves he does not expect AI to fully replace doctors.

Asked what in his unusual childhood education could be scaled with AI, Abraham points first to the availability of online resources, which he says would not have been possible decades ago. He adds that most education systems implicitly assume everyone learns at the same rate, which caused him problems as a fast learner. AI could provide personalized programs that meet each student where they are, at whatever pace.

He cautions that the obstacle is social rather than technical, because personalized education implies an overhaul of classrooms and the way schools work. How society reacts to and implements AI is a separate hard question in any field.

On whether he would get a Neuralink once millions of people had one, he says he would love the idea in theory but worries about security, since he does not want anyone hacking his brain. Setting that aside, he would gladly augment himself with a brain-computer interface.

On his fears, Abraham names the risk that AI could make society more polarized, that biases could be amplified, and that unreliable AI tools could lead to dangerous situations. He treats these as real concerns rather than distant ones.

On his hopes, he leads with medical applications and patient care, then education, where AI could let people learn whatever they want. He also thinks AI can bring people together, noting that even primitive tools like translation already connect us to other cultures.

He is especially interested in AI augmenting creativity, helping people bring ideas into the real world, perhaps eventually through mind reading that carries an idea straight from imagination to something physical. He ends by calling it a nuanced topic where benefits and risks go together, and Labenz wishes him a happy 20th birthday.