Blur need no longer be a dead end. From GANscan and Blurryscope to his latest system, SCANIMUS, Michael Fanous explores how controlled motion could make microscopy more accessible without compromising the evidence.

I believe the microscope is among our most underestimated engineering achievements. Coupled with motorised control, continuous scanning and taskspecific AI (and, eventually, models periodically refined on independently validated data) this familiar piece of scientific furniture could become one of the laboratory’s most dazzling systems.

Watching it interrogate a specimen rapidly can be striking. Deep learning lets engineers explore deliberate, measurable compromises, such as accepting motion blur to increase acquisition speed. But any speed gain is valuable only if reconstruction fidelity or task performance is independently validated against high-quality reference images or expert-verified labels and meets predefined, task-specific thresholds.

That principle was explored in GANscan and BlurryScope and now informs the development of SCANIMUS. The goal is not to eliminate end-user interpretation, but to place the relevant regions, underlying measurements and supporting context before the expert at the final evaluation.

The engineering question follows: can laboratories relax some traditional hardware constraints, accept imperfect raw data and still recover information useful for a defined task? Sometimes they can – but only if we are precise about the task, the information retained by the measurement and the limits of the algorithm.

Many scanning microscopes use a “stop- and-stare” sequence: move the specimen, stop, allow vibrations to settle and capture an image. This protects sharpness but adds time and mechanical demands. Continuous scanning removes much of that overhead and can increase throughput. Some systems already use specialised cameras or synchronised illumination. Our work explored accepting controlled motion blur with conventional camera hardware and measuring how much blur could be tolerated before reconstruction fidelity or task performance deteriorated.

Motion blur is normally a problem because the specimen continues moving during the exposure. A feature is spread across the distance it travels.

In BlurryScope’s reported configuration, a specimen speed of five millimetres per second and an exposure of 7.8 milliseconds produced a blur distance of roughly 39 micrometres. Closely spaced structures can appear to merge, and boundaries can become difficult to judge. A shorter exposure reduces blur but also collects less light unless illumination changes. The trade-off is physical and measurable [1].

In GANscan, work I led with the late Professor Gabriel Popescu at the University of Illinois Urbana–Champaign, we trained a neural network on corresponding motion-blurred and sharp micrographs. The microscope recorded video while the specimen moved continuously, and the model learned to estimate a sharp image from each blurred input. In the reported experimental system, this achieved up to 30 times the acquisi- tion throughput of stop-and-stare imaging. It let us treat acquisition and reconstruction as parts of the same instrument [2].

Does such a model recover information from the blurred image, or generate a plausible representation from its training data? The honest answer is that it draws on both the recorded signal and learned patterns. When enough information survives, the measurement constrains the reconstruction. Where fine detail is ambiguous or lost, the model cannot know with certainty which structure was present; different original structures may have produced a similar blurred image. Its output is then an estimate, not a direct measurement of ground truth.

That distinction creates the central safety risk. A convincing reconstruction can introduce a feature, alter its shape or conceal something real – including a diagnostically relevant feature. Visual attractiveness is therefore an inadequate test of fidelity. Reconstructions should be compared with independently acquired sharp images and tested on specimens and patients excluded from training.

Our work explored accepting controlled motion blur with conventional camera hardware and measuring how much blur could be tolerated before reconstruction fidelity or task performance deteriorated

Biologically important measurements must remain reliable across changes in focus, staining and motion. Raw data should be preserved, uncertainty and failure states made visible, and conventional sharp reacquisition kept available. A careful instrument must recognise when stopping the stage remains necessary [3].

BlurryScope, developed with colleagues in the Bio- and Nano-Photonics Laboratory at UCLA, explored a related but importantly different route: predicting a defined tissue score directly from motion-blurred images rather than first reconstructing a sharp image. Our 2025 study tested HER2 scoring, a biomarker task in breast tissue, on a blinded set of 284 unique patient cores. It reported 79.3% accuracy across four HER2 classes and 89.7% when grouping 0/1+ versus 2+/3+. This was a research proof of concept within that specific setting.

Direct classification and image reconstruction answer different questions; success at one does not establish the other, and neither alone validates general diagnostic use.

This distinction also shapes the workflow we want to build. An experienced microscopist surveys a specimen, selects an area, changes magnification and refocuses. An automated instrument could support that sequence while preserving the ability to look again. That thinking informs SCANIMUS, a compact, 35-centimetre-tall research-use-only system we are developing at scanimus.ai. Coupled to actuators and computation, a microscope points towards an intelligent imaging robot whose movement and acquisition are designed to respond to the investigation. Yet an impressive reconstruction means little unless the instrument is transparent and practical to use.

The same question reaches beyond microscopy. Could a spectrometer tolerate lower optical resolution if a validated model could still estimate specific constituents? Could another instrument collect fewer measurements for a tightly defined task? Perhaps – but each application must be tested separately. Savings in optics or mechanics must be weighed against calibration, representative data, computing, maintenance and continuing quality control. Complexity does not disappear; it changes location and may acquire a different cost.

Many scanning microscopes use a “stop- and-stare” sequence: move the specimen, stop, allow vibrations to settle and capture an image. This protects sharpness but adds time and mechanical demands

There may also be a consequence for demand. In The Coal Question, published in 1865, William Stanley Jevons argued that using coal more efficiently could expand its uses enough to increase total consumption. The Jevons paradox describes cases in which growth in use outweighs the resource savings per application; efficiency alone does not guarantee that outcome [4].

For microscopy, I see an analogy worth testing. If obtaining and analysing images becomes easier and cheaper, laboratories may ask more questions of specimens they already possess. Archived tissue could become accessible to additional research, and existing slides could yield further measurements. My hope is for more information from each specimen already collected. Whether demand expands will depend on scientific value, budgets and the capacity to interpret the results.

Better access to imaging could allow many more laboratories to extract useful information from specimens. The engineering challenge is to make instruments easier to use and measurements easier to obtain, while keeping every inference accountable to the evidence.

References:

1 Fanous MJ et al. “BlurryScope enables compact, cost-effective scanning microscopy for HER2 scoring using deep learning on blurry images.” npj Digital Medicine 8, 506 (2025). DOI: 10.1038/s41746-025-01882-x

2 Fanous MJ and Popescu G. “GANscan: continuous scanning microscopy using deep learning deblurring.” Light: Science & Applications 11, 265 (2022). DOI: 10.1038/s41377-022-00952-z

3 Belthangady C and Royer LA. “Applications, promises, and pitfalls of deep learning for fluorescence image reconstruction.” Nature Methods 16, 1215–1225 (2019). DOI: 10.1038/s41592-019-0458-z

4 Jevons WS. The Coal Question. London: Macmillan; 1865