The research, conducted by Berkay Çağdaş of the Department of Nuclear Medicine at Afyonkarahisar State Hospital and Ege University’s Institute of Nuclear Sciences, and Elif Kardelen Çağdaş of the hospital’s Department of Pathology and Ankara University’s Graduate School of Biotechnology, set out to determine whether regional differences in immune and stromal composition inside breast tumors correspond to measurable differences in how the tumor consumes glucose on PET imaging. The premise is biologically plausible: tumors infiltrated by lymphocytes often behave differently from immune-cold tumors, and fibroblast-like stromal cells can alter metabolism, perfusion, and tissue architecture in ways that might plausibly change fluorodeoxyglucose uptake.

The study enrolled 61 women with invasive breast carcinoma who underwent baseline 18F-FDG PET/CT before treatment. Rather than treating each tumor as a single homogeneous blob, the team took a deliberately regional approach. A pathologist, blinded to all PET findings, annotated ten separate intratumoral regions on whole-slide histopathology images, each at least one square millimeter in area. QuPath, an open-access machine-learning platform for digital pathology, then quantified two key features within each region: the density of stromal tumor-infiltrating lymphocytes, the immune cells that gather in the connective tissue surrounding cancer cells, and the density of fibroblast-like stromal cells, spindle-shaped cells identified morphologically on routine hematoxylin and eosin stains.

From those ten regions, the investigators retained the ones with the numerically highest and lowest lymphocyte density for paired analysis, an approach designed to capture the observed range of immune-stromal composition within each tumor. The authors are careful to note that these paired regions represent the extremes of what was measured, not a comprehensive spatial map of every tumor. On the imaging side, the team extracted a rich panel of 30 PET features from each patient’s scan. These included the familiar intensity measures such as maximum, mean, and peak standardized uptake values, volumetric measures including metabolic tumor volume and total lesion glycolysis, texture descriptors derived from gray-level co-occurrence, run-length, and size-zone matrices, and dedicated heterogeneity indices, including the HI2 and HI3 indices and a slope calculated across metabolic volume thresholds from 30 to 50 percent of maximum uptake.

The statistical framework was deliberately conservative. The team ran Spearman correlations between nine immune-stromal endpoints and all 30 PET features, generating 270 individual tests, and then applied Benjamini-Hochberg correction to control the false discovery rate, a standard safeguard against being fooled by chance when many comparisons are made simultaneously. The results were sobering but informative. Thirty-eight of the 270 correlations reached nominal significance at the conventional p-value threshold of 0.05, yet not a single one survived the false discovery rate correction, whether the correction was applied within each endpoint family or across the entire study. The smallest corrected q-value within an endpoint was 0.077, and study-wide it was 0.33, both well above any threshold that would conventionally be called significant.

The strongest raw association was, by the authors’ own framing, weak. The density of stromal tumor-infiltrating lymphocytes in the TIL-rich region correlated with the HI2 heterogeneity index at a Spearman coefficient of 0.379 in absolute value, with a 95 percent confidence interval spanning 0.126 to 0.586 and a nominal p-value of 0.0026. Beyond that single headline number, a coherent pattern emerged in the direction of the nominal findings. Lymphocyte density correlated inversely with metabolic tumor volume, with coefficients ranging from negative 0.29 to negative 0.35, and positively with the HI2 heterogeneity measure. The ratio of lymphocytes to fibroblast-like stromal cells correlated with SUV-based parameters, the largest being a coefficient of 0.299 with mean standardized uptake value. Fibroblast density on its own, and the metrics describing regional differences between the paired areas, produced weaker and less consistent signals.

One of the study’s most technically rigorous contributions is its validation of the pathology measurement itself. Because fibroblast-like stromal cells were defined morphologically on routine H&E stains, the authors cross-checked their machine-learning quantification against alpha-smooth muscle actin immunohistochemistry, a stain that highlights activated fibroblasts and myofibroblasts, in the same 122 paired regions. The agreement was remarkably tight: a correlation of 0.992 and a Lin concordance correlation coefficient of 0.996. When the entire correlation analysis was repeated using the immunohistochemistry-based stromal density instead of the H&E-based one, the profile of results was reproduced almost exactly, with correlations between the two analyses ranging from 0.94 to 0.996. This means the null finding is not easily explained by measurement error in the pathology pipeline.

The team also stress-tested the associations that did reach nominal significance. When extreme lymphocyte densities or the smallest annotated regions were excluded from the analysis, between 9 and 24 of the original 38 nominal associations persisted, indicating that some of the signals were fragile and dependent on particular data points. After statistical adjustment for clinical covariates including T stage, Nottingham histologic score, Ki-67 proliferation index, and molecular subtype, only 13 of the 38 associations remained nominally significant. The picture that emerges is one of faint, directionally interesting signals that dissolve under the weight of proper multiple-comparison correction, exactly the pattern one expects when exploring a genuinely uncertain hypothesis with a modest sample size.

Why does this matter? PET radiomics has generated enormous enthusiasm in oncology, with hundreds of studies proposing imaging texture features as noninvasive surrogates for everything from gene expression to immune infiltration. The promise is seductive: if a scan could reveal the immune landscape of a tumor without a biopsy, clinicians could select patients for immunotherapy, track treatment response, and monitor clonal evolution noninvasively. But the field has been repeatedly criticized for underpowered studies, lax statistical correction, and features that fail to replicate. This study is a refreshing counterexample. By pre-specifying a conservative statistical framework, validating its pathology measurements against an independent stain, and openly reporting that nothing survived correction, the authors have modeled the kind of intellectual honesty that imaging science needs more of.

The authors themselves are unambiguous about the limits of their work. This is a hypothesis-generating, exploratory, single-institution study of 61 patients, and the paired-region design captures only the extremes of immune-stromal composition rather than a full spatial map. They state plainly that the findings require independent validation in larger, multi-institutional cohorts before any biological or clinical interpretation can be drawn. That caution is warranted, but so is the value of the attempt. The study demonstrates a feasible, reproducible workflow for integrating region-level digital pathology with PET radiomics in breast cancer, and it establishes a realistic benchmark for how weak the true links between tissue microenvironment and metabolic imaging may actually be. If future larger studies confirm even the modest directional trends seen here, immune-rich tumors may one day be recognizable on PET as smaller, more heterogeneous glucose-avid lesions. For now, the honest answer to the question of whether a PET scan can see a tumor’s immune landscape is: perhaps faintly, and not yet reliably.

Subject of Research: Associations between regional immune-stromal tumor features and 18F-FDG PET metabolic heterogeneity in breast cancer

Article Title: Exploratory associations between regional immune-stromal features and 18F-FDG PET-derived metabolic heterogeneity in breast cancer

Article References: Çağdaş, B., & Çağdaş, E. K. (2026). Exploratory associations between regional immune-stromal features and 18F-FDG PET-derived metabolic heterogeneity in breast cancer. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02789-z

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02789-z

Keywords: breast cancer, PET imaging, radiomics, tumor microenvironment, tumor-infiltrating lymphocytes, digital pathology, metabolic heterogeneity, 18F-FDG, QuPath, fibroblasts, false discovery rate, machine learning

The problem the researchers set out to solve is familiar to anyone who has ever photographed a Petri dish on a laboratory bench. Petri dishes and agar surfaces have reflective and optical properties that conspire against clean imaging. Photographs taken under routine room lighting or with handheld mobile phones are frequently marred by glare, reflections, shadows, and uneven illumination. These artifacts are not merely cosmetic. They substantially alter the pixel intensity distributions across agar plate images, introducing nonbiological variation, or noise, that presents serious challenges to automated analysis. As machine learning tools increasingly support automated colony counting, morphological comparison, and quantitative assessment of media color changes that signal metabolic processes such as pH shifts or hemolysis, the quality of the image captured at the point of acquisition ultimately determines the reliability of everything that follows.

The lightbox itself is an elegant exercise in accessible engineering. The initial model was developed in Tinkercad and refined in Autodesk Fusion, and the device incorporates an enclosed imaging chamber, an internal platform for positioning agar plates, removable diffusion and light-blocking components, and integrated light-emitting diode lighting. The chamber measures approximately 220.7 millimeters deep by 200 millimeters high by 200 millimeters wide and accommodates standard 90-millimeter Petri dishes, with a modular plate-support insert that can be modified for alternative plate sizes. An aperture at the top allows a smartphone camera to be reproducibly positioned directly above the plate, standardizing imaging distance, camera angle, and plate position between photographs.

Two independently configurable LED strips provide illumination from above and below the agar plate. The lower source delivers transmitted illumination through the agar, while the upper source provides reflected illumination to assist visualization of colony morphology, pigmentation, and media-associated changes. A removable one-millimeter diffuser plate, printed in white, softens the direct LED illumination from below, and a removable black blocker plate isolates individual lighting configurations. All printed components were fabricated in white polylactic acid filament, chosen because it is inexpensive, widely available, and compatible with most desktop 3D printers, while the white color enhances internal light reflection and promotes more uniform illumination. The main enclosure required approximately 700 to 900 grams of filament, corresponding to an estimated material cost of about AU$14 to AU$36, and the two LED light strips added roughly AU$15 each. Design files, assembly notes, and supporting workflow materials are freely available through an open-access repository to support reuse, modification, and local fabrication.

The performance evaluation compared three imaging conditions across five uncultured agar types: bile aesculin, nutrient, horse blood, mannitol salt, and MacConkey agar. Photographs were captured with the same iPhone 13 Pro under identical camera settings with flash disabled, either on a traditional black background, inside the lightbox with top lighting only, or inside the lightbox with combined top and bottom lighting. Analysis using Fiji/ImageJ examined 3D surface plots and pixel color histograms, in which reflective artifacts appeared as sharp peaks and uneven illumination showed up as greater dispersion of pixel colors. Across all plate conditions, lightbox imaging produced more even pixel color distributions than traditional photography, which routinely showed reflections from the phone itself and from the biosafety cabinet filter screen.

The bottom-lit configuration proved especially significant for hemolysis assessment. The American Society for Microbiology recommends that hemolytic reactions on blood agar be assessed with the light source positioned behind the plate to enhance visualization of zones surrounding colonies that may otherwise be faint. In cultured plate tests, the combined top and bottom lighting produced greater color and contrast differentiation between alpha, beta, and gamma hemolysis patterns, improving visual distinction between hemolytic classes. Transmitted illumination through the agar also improved visibility of small or low-contrast colonies that were less apparent under traditional imaging, while reducing reflective artifacts across both uncolonized regions and individual colonies.

To explore whether these improvements matter for machine learning, the team turned to QuPath, a freely available image analysis platform that lets users develop trainable pixel classifiers without coding expertise. Fourteen growth-positive horse blood agar plates, prepared from swabs of hospital staff mobile phones in an intensive care unit environment, were imaged under both conditions. Separate Random Trees pixel classifiers were trained for traditional and lightbox images of the same plates, with identical numbers of training annotations, allowing the effect of image acquisition conditions to be compared while holding the underlying colony distribution constant. The results were striking. For a representative matched plate, the classifier trained on the traditional image generated 159 detections, with visual review identifying false positives caused by printed plate text being detected as colony-like objects, uncolonized agar regions classified as colonies, and individual colonies fragmenting into multiple detections. The classifier trained on the corresponding lightbox image generated just 76 detections and showed fewer artifact-related errors.

An exploratory comparison with manual counting provided further support for the workflow’s potential. Across the 14 growth-positive plates, manual assessment identified 1014 colonies, compared with 992 from the semiautomated QuPath-assisted workflow following human review. The counts showed a strong linear association, with a Pearson correlation of 0.999 and a mean absolute error of just 2.4 colonies per plate. The authors are careful to frame these results as exploratory rather than validated, since classifier refinement and evaluation used the same images without an independent holdout data set. Larger independent data sets incorporating a broader range of organisms, colony densities, and imaging conditions will be needed before the workflow can be considered validated.

The study also highlighted stubborn artifacts that no lighting scheme can eliminate. Factory-printed text on the underside of agar plates, carrying agar type, batch, and expiry details, remained visible under all conditions and was particularly problematic on opaque media such as horse blood agar, where it appeared as darker, poorly defined regions resembling colonies or hemolytic changes. The researchers suggest training classifiers to recognize printed text as a separate image class, though this increases the annotation burden. The circular rim on the plate base similarly introduced edge-related artifact from light refraction and geometry changes, which can be managed by cropping to the agar area within the bottom rim.

Beyond the computational applications, the authors see real promise for microbiology education. Standardized photography could support more consistent image sets for practical teaching, assessment, and online learning resources, while giving students hands-on experience with digital image analysis and machine learning concepts. By comparing images captured under controlled and uncontrolled conditions, students can directly observe how imaging artifacts and data quality shape computational outputs, a principle that resonates across microbiology, histopathology, and radiology. The device also builds on earlier open-source work, including a prior imaging box assembled from cardboard, wood, and fabric, adding reusable 3D-printed components, a fixed camera aperture, and independently configurable dual illumination. Looking ahead, the authors note that alternative printing materials such as polyethylene terephthalate glycol may better withstand repeated disinfection in routine laboratory use, and that any change in material or surface finish should be reevaluated because it could alter illumination conditions and the performance of models trained under them. Future work should also assess color fidelity with standardized references and explore whether ambient light combined with smartphone night mode might offer an even simpler path to consistent imaging.

Subject of Research: A low-cost 3D-printed lightbox for standardized smartphone imaging of agar plates to support machine learning-assisted microbiological analysis.

Article Title: Improving Agar Plate Image Quality Using a Low‐Cost 3D‐Printed Lightbox: A Proof‐of‐Concept Evaluation for Machine Learning–Assisted Analysis

Article References: Improving Agar Plate Image Quality Using a Low‐Cost 3D‐Printed Lightbox: A Proof‐of‐Concept Evaluation for Machine Learning–Assisted Analysis. (n.d.). https://doi.org/10.1002/mbo3.70399

Image Credits: AI Generated

DOI: 10.1002/mbo3.70399

Keywords: 3D-printed lightbox, agar plate imaging, machine learning, colony counting, QuPath, pixel classification, hemolysis, smartphone microscopy, open-source hardware, microbiology education, image quality, automated colony counting