The core of the method is a two-stage pipeline. First, photographs of rockfill material—whether spread on a laboratory bench or compacted in layers on a dam site—are fed into Cellpose, a generalist deep learning algorithm originally developed for segmenting individual cells in microscopy images. Cellpose’s neural network learns to delineate the boundaries of touching, overlapping, irregularly shaped objects, which makes it unusually well suited to the visual chaos of a pile of quarried rock. The algorithm produces a high-precision segmentation mask in which every individual particle is separated from its neighbors, solving the classic bottleneck that has limited automated rock image analysis: stones that touch or partially overlap are notoriously difficult for conventional thresholding techniques to separate.

Once Cellpose has produced the segmented image, the researchers turn to the MorphoLibJ plugin of ImageJ, a mature mathematical morphology library, for post-processing and quantitative computation. From each segmented particle, the software extracts geometric features including area, particle size, and aspect ratio. Aggregated over the hundreds or thousands of stones visible in a single image, these features yield the particle size distribution curve—one of the most fundamental descriptors of any granular construction material—without the sieving, weighing, and labor that traditional gradation testing demands. The team reports that the ImageJ-Cellpose method could accurately determine both the particle size distribution curve and the porosity of rockfill materials, results that were validated against conventional screening and porosity experiments conducted in the laboratory.

Beyond reproducing established metrics, the study introduces new shape parameters for rockfill material, including what the authors call the long-short ratio and the area ratio. Shape matters in geotechnical engineering more than casual observers might expect: angular, elongated, and flat particles interlock differently than rounded, equidimensional ones, influencing shear strength, particle breakage under load, and the volumetric behavior of the compacted mass. By statistically analyzing the distribution of these new shape descriptors across particle populations, the researchers offer a quantitative framework for describing rockfill morphology that goes beyond simple size gradation, opening a path toward correlating shape statistics directly with mechanical performance.

The critical question for any laboratory method is whether it survives contact with the real world. To answer it, the team partnered with a pumped-storage power station, where layered rolling experiments were carried out during actual rockfill placement. As the material was compacted by rollers in successive passes, images were collected at each stage and processed through the intelligent recognition pipeline. The results showed that the method could effectively trace how the particle size distribution and porosity of the rockfill evolved with each rolling pass—precisely the kind of real-time feedback that compaction quality control has historically lacked. Encouragingly, the trends identified by the image-based approach aligned well with those derived from the pit measuring method, the traditional but destructive and labor-intensive standard for evaluating compaction in the field.

The significance of this validation lies in what it could change about dam construction practice. Rockfill dams are among the largest man-made structures on Earth, and their long-term safety depends on how well the fill is compacted during construction; poor compaction leads to post-construction settlement, cracking of upstream concrete face slabs, and heightened vulnerability to overtopping failure and seismic loading. Current quality control typically relies on periodic pit sampling, in which crews excavate a hole, weigh the removed material, and measure density—a slow process that samples only a tiny fraction of the dam’s volume. A camera, a laptop running Cellpose and ImageJ, and the proposed analytical workflow could in principle assess compaction indicators continuously and across the entire working surface, converting construction monitoring from intermittent spot checks into comprehensive surveillance.

Technically, the study is notable for transplanting tools from an unexpected domain. Cellpose was created by biologists to solve the problem of segmenting cells in diverse image types without retraining a network for each new dataset, and its generalist design proved transferable to geological imagery. ImageJ, likewise, grew out of biomedical imaging but ships with general-purpose morphology operators that apply equally well to mineral particles. By connecting a state-of-the-art segmentation network to a battle-tested analysis library, the Xi’an team sidestepped the need to build bespoke deep learning infrastructure or large annotated rockfill datasets from scratch, lowering the barrier to adoption for engineering firms and construction supervisors who cannot maintain their own machine learning teams.

The method also complements a growing body of work on computer vision in geotechnical monitoring. Prior studies have used video image recognition to detect rockfill gradation, deep learning segmentation to calculate gradation from photographs, and instance segmentation to predict gradation for engineering projects. What distinguishes the new approach is its end-to-end coverage: rather than estimating only particle sizes, it simultaneously delivers size distribution, individual particle shape metrics through newly proposed parameters, and porosity, all from ordinary images, and it does so in a form validated against both sieving tests in the laboratory and pit measurements on an active construction site. The authors argue that this makes the technique suitable for real-world engineering applications and a robust means of ascertaining particle size information during the construction of rockfill dams.

Looking forward, the research points toward a swift and precise evaluation method for the compaction quality of rockfill material—a capability the authors describe as paramount for the comprehensive understanding and analysis of the safety conditions of rockfill dams and similar projects. If deployed at scale, the ImageJ-Cellpose workflow could give dam engineers a continuous, quantitative picture of how each roller pass is reshaping the granular skeleton of an embankment, catch gradation or compaction anomalies before they are buried under the next lift of rock, and feed high-quality geometric data into the numerical models used to predict settlement, seepage, and seismic response. In an era when aging dam infrastructure faces intensifying climate and seismic stresses, the idea that a deep learning algorithm designed for counting cells can help safeguard structures holding back billions of liters of water is a striking demonstration of how tools migrate across scientific boundaries—and of how much of modern construction safety may soon rest on image recognition pipelines that work quietly, frame by frame, on the rocks beneath the rollers.

Subject of Research: Intelligent image-based geometric characterization of rockfill materials using deep learning segmentation for dam compaction analysis

Article Title: Intelligent geometric characterization of rockfill materials using ImageJ-Cellpose: a novel approach for accurate morphological analysis

Article References: Ma, C., Zhou, C., Hou, Y., & Cheng, L. (2026). Intelligent geometric characterization of rockfill materials using ImageJ-Cellpose: a novel approach for accurate morphological analysis. Cluster Computing, 29(13), Article 779. https://doi.org/10.1007/s10586-026-06480-4

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06480-4

Keywords: rockfill materials, deep learning, Cellpose, ImageJ, image segmentation, particle size distribution, porosity, particle shape, compaction quality, rockfill dams, image recognition, geotechnical engineering

The new workflow, described in Physiological Reports, measures five key endpoints simultaneously: the number of myonuclei per muscle cell, the number of Pax7-positive satellite cells per cell, the cross-sectional area of each fiber, the perimeter of each fiber, and the fiber’s type, classifying it as slow-twitch type I or fast-twitch type II. According to the authors, this is the first published pipeline to combine all five measurements in a single workflow compatible with widely available widefield microscopy, rather than requiring specialized imaging platforms. Each of these endpoints carries real biological weight. Satellite cells are the resident stem cells of skeletal muscle, and their expansion and integration into fibers as new myonuclei underpins theories such as the myonuclear domain hypothesis, which holds that each nucleus governs a finite volume of cytoplasm. Emerging evidence even suggests the domain may track fiber perimeter more closely than cross-sectional area, making accurate perimeter measurement newly important.

The pipeline begins before any computation, with careful tissue handling and staining. Human biopsies were taken from the vastus lateralis, the lateral thigh muscle that is a standard site in exercise physiology studies. Tissue was embedded in optimal cutting temperature compound, flash-frozen in liquid-nitrogen-cooled isopentane, and sectioned at seven micrometers on a cryostat. Immunohistochemistry then labeled three targets: laminin, a protein that outlines the sarcolemma surrounding each fiber; type I myosin heavy chain, which marks slow fibers; and Pax7, the canonical marker of quiescent and activated satellite cells. Slides were imaged at twenty-fold magnification on a Leica widefield microscope across four channels—DAPI for all nuclei, green fluorescence for Pax7, yellow for myosin heavy chain I, and far-red for laminin—producing the raw material for analysis.

A crucial early stage is manual image cleaning in Fiji/ImageJ. Rather than treating every pixel as trustworthy data, the protocol calls for brightness and contrast optimization on each channel, followed by manual removal of artifacts such as tissue folds, blood vessels, blurred or overlapping laminin signal, and out-of-focus nuclei. The authors argue this step, though the most time-consuming part of the workflow, is essential: it strips out the noise that would otherwise generate false positives downstream and brings the analyzed regions into line with what a careful human analyst would measure. Cleaned images are then exported with a strict naming convention that encodes subject, timepoint, leg, and set, enabling fully automated batch processing.

The computational core pairs two open-source tools. CellProfiler, the modular image-analysis platform maintained at the Broad Institute, orchestrates segmentation and measurement, while Cellpose, a machine-learning model developed at HHMI’s Janelia Research Campus, performs deep-learning-based detection of cellular boundaries. Fiber segmentation uses Cellpose’s cyto2 model on the laminin channel to trace sarcolemmal borders. Nuclear segmentation is trickier, because muscle cross-sections contain nuclei from many cell types, and the default Cellpose nuclei model is highly sensitive. The pipeline handles this with an iterative subtraction strategy: a Gaussian filter smooths the DAPI signal, an initial Cellpose run identifies candidate nuclei, background intensity statistics are computed, the seventeenth percentile of pixel intensity is subtracted to suppress non-specific fluorescence, and a refined second run produces the final nuclear masks.

Notably, deep learning proved less suitable for the rarest target. Pax7-positive satellite cells are sparse—on average only about one per ten fibers in the validation dataset—and their signal is faint against the sarcoplasmic background. For these, the pipeline falls back on CellProfiler’s classical IdentifyPrimaryObjects module, using Otsu thresholding with size, circularity, and intensity filters tuned to the amplified Pax7 signal. The authors note the modular design means this component could be swapped to detect other sparse cell populations; they have already adapted the same pipeline to identify macrophages and atrophied fibers in an elderly clinical cohort.

A custom Python script built on OpenCV, pandas, and NumPy then fuses the separate segmentation masks into a single biological dataset. Each fiber mask is decomposed by its unique grayscale label, and every nucleus or satellite cell is assigned to the fiber with which it shares the greatest bitwise pixel overlap—a ‘greatest overlap’ rule that resolves objects straddling two fibers. Fiber area and perimeter are computed in physical units using a pixel-to-micron scaling factor, and fiber type is assigned automatically by measuring mean type I myosin heavy chain fluorescence within each fiber mask after rolling-ball background subtraction. The intensity threshold separating type I from type II fibers, set at a pixel value of twenty, was derived empirically from an analysis of 3,722 cells across twelve human biopsies. Output is a unified Excel workbook listing fiber ID, area, perimeter, nuclear count, satellite cell count, and typing for every cell, alongside diagnostic overlay images that allow researchers to audit the automated calls visually.

Validation against traditional manual analysis is where the pipeline earns its credentials. The team compared both methods head-to-head on 310 matched cells from a human biopsy. For myonuclei per cell, manual counting averaged 2.471 nuclei versus 2.474 for the pipeline, a difference of just -0.003 nuclei, with no statistically significant difference by paired t-test. Satellite cell counts likewise agreed, averaging 0.110 manually versus 0.119 automatically. Fiber typing was the strongest performer: the pipeline classified 49 type I and 261 type II fibers against the manual count of 48 and 262, with a correlation of r = 0.939 between methods. Cross-sectional area and perimeter correlated even more highly, at r = 0.959 and r = 0.935 respectively, though with systematic offsets that the authors explain candidly.

Those offsets trace to boundary conventions rather than measurement failure. Manual outlines were drawn with a ten-pixel brush slightly inside the laminin border, while the automated segmentation places its one-pixel outline on or marginally outside it. Because area and perimeter are computed from within the drawn boundary, this small difference inflates automated values—the pipeline averaged 3,532 versus 3,005 square micrometers for area, and 264 versus 249 micrometers for perimeter. Similar discrepancies between manual and automated fiber outlines have been reported before, and the authors stress that users simply need to apply consistent methodology. Other limitations include the mandatory image-cleaning step, strict file-naming requirements, and occasional detection failures on very sparse or dim fields of view, for which the authors suggest alternative Cellpose detection modes or the newer Cellpose3.

The practical payoff is throughput. Published power calculations indicate that comparisons of satellite cells per fiber require at least fifty type I and seventy-five type II fibers, and many labs have settled on a minimum of 150 cells per biopsy as a working threshold. A single human biopsy can yield well over a thousand fibers, and the new pipeline can process all of them, dramatically increasing sample representation while reducing both analysis time and operator dependence. Every cell receives a unique identifier, every segmentation decision is saved as a reviewable overlay, and the modular CellProfiler architecture invites adaptation to new antibodies, tissues, and imaging conditions. As interest grows in satellite cell dynamics across the human lifespan—from athletic performance to sarcopenia and rehabilitation—the authors argue this workflow offers muscle biologists a scalable, transparent, and reproducible foundation for turning microscope images into meaningful numbers.

Subject of Research: A semi-automated microscopy pipeline for quantifying satellite cells, myonuclei, and muscle fiber morphology in human skeletal muscle cross-sections

Article Title: A semi‐automated pipeline for quantitation of Pax7+, myonuclei, and cross‐sectional area by fiber type

Article References: Megowan, H. G., Luu, M., Shuaib, A., Augienello, K. B., Fries, A. C., Searcy, J., & Dreyer, H. C. (2026). A semi‐automated pipeline for quantitation of Pax7+, myonuclei, and cross‐sectional area by fiber type. Physiological Reports, 14(18), Article e71097. https://doi.org/10.14814/phy2.71097

Image Credits: AI Generated

DOI: 10.14814/phy2.71097

Keywords: skeletal muscle, satellite cells, Pax7, myonuclei, cross-sectional area, fiber type, image analysis, CellProfiler, Cellpose, immunohistochemistry, muscle biology, widefield microscopy