Amnestic mild cognitive impairment, or aMCI, is widely regarded as the prodromal phase of Alzheimer’s disease, a stage where memory complaints are measurable but daily function is largely preserved. Not everyone in this stage progresses, however, and clinicians have long sought tools to separate those carrying underlying beta-amyloid pathology from those whose memory problems stem from other causes. Amyloid positron emission tomography answers that question by using fluorine-18-labelled radiotracers that bind to beta-amyloid deposits in the cortex, making the molecular signature of the disease visible in a living brain. Patients whose scans show pathological cortical retention are known to be at substantially higher risk of clinical decline, which is why amyloid PET has become central to diagnosis and, increasingly, to the selection of patients for anti-amyloid therapies.

Yet the way a scan is interpreted has always been a point of debate. Visual reading, the current clinical standard, depends on the trained eye of a nuclear medicine physician comparing grey and white matter contrast against tracer-specific criteria. It is fast and well validated, but it is inherently subjective, and borderline scans with low or intermediate tracer retention can divide even experienced readers. To reduce that subjectivity, imaging platforms now offer semi-quantitative support: standardized uptake value ratios, or SUVRs, that condense tracer signal in composite cortical regions into a single number compared against validated thresholds, and z-score maps that highlight voxels deviating from a normative database of age-matched healthy controls. The updated Appropriate Use Criteria from the Alzheimer’s Association and the Society of Nuclear Medicine and Molecular Imaging endorse these tools as complements to visual interpretation in selected scenarios, but whether they genuinely improve reproducibility or prognostic power beyond expert reading has remained uncertain.

The Navarra team set out to answer that question directly in a routine memory clinic setting. Between October 2013 and March 2021, they enrolled 145 patients diagnosed with aMCI after comprehensive neurological, neuropsychological, laboratory and MRI workups, excluding anyone over 85, those with depression, abnormal laboratory findings, or alternative brain injuries such as lacunar infarcts. Three different fluorinated amyloid tracers were used in clinical practice: florbetapir in 79 patients, flutemetamol in 46, and florbetaben in 20, all acquired on the same Biograph mCT PET/CT scanner with time-of-flight and point-spread-function reconstruction. Patients were then followed clinically until January 2022, with progression defined by the National Institute on Ageing–Alzheimer’s Association 2011 criteria for Alzheimer’s disease dementia.

Two certified nuclear medicine physicians, one with more than 30 years of experience and one with five, independently read every scan twice: first with conventional structured visual interpretation, and then with assistance from the syngo.via MI Neuro Cortical Analysis software, which overlays voxel-wise z-score statistics and SUVR values onto the images. As a third method, scans were classified purely by SUVR thresholds validated in prior literature, above 1.17 for florbetapir, above 1.35 for florbetaben and above 0.60 for flutemetamol. Agreement between and within readers was quantified with Cohen’s kappa, a statistic in which values above 0.81 conventionally signal excellent agreement.

The reliability results were striking. Inter-observer agreement reached a kappa of 0.877 for structured visual reading and 0.904 for assisted interpretation, while intra-observer consistency between the two approaches was 0.860 for the senior reader and 0.892 for the junior one. When consensus ratings were established with the assisted method, agreement climbed to 0.936, near the statistical ceiling. Reader 1 classified 94 scans as amyloid-positive on visual reading and 101 with assistance; Reader 2 classified 96 and 97 respectively. In other words, experienced readers agreed with themselves and with each other at a level that leaves very little room for chance, confirming that standardized visual interpretation of amyloid PET is far more reproducible than its subjective reputation might suggest.

Reproducibility, however, is only half the story; the real question is whether a positive scan predicts what happens next. Using Cox proportional hazards models adjusted for age, sex and education, the researchers found that amyloid positivity significantly increased the risk of conversion to Alzheimer’s dementia under all three interpretation methods. Assisted visual interpretation produced the strongest signal: PET-positive patients had a 3.23-fold higher risk of progression than PET-negative patients, with a hazard ratio of 3.23 and a 95 percent confidence interval of 1.94 to 5.38. SUVR quantification yielded a hazard ratio of 2.19, and conventional visual reading, 1.83. Kaplan–Meier survival curves separated significantly between positive and negative groups for every method, with log-rank tests all significant below the 0.001 level. Over the follow-up period, 75 patients converted to Alzheimer’s dementia, while 41 remained clinically stable.

The most intriguing findings came from the scans where readers disagreed. Eighteen of the 145 studies, 12.4 percent, showed discordance across at least one interpretative condition, and in this small, exploratory subgroup the hierarchy of methods flipped. When clinical progression was used as the pragmatic reference standard, conventional visual reading outperformed assisted interpretation for both readers: the senior reader correctly classified 9 of 18 discordant cases visually versus only 2 with assistance, and SUVR binary classification matched clinical outcome in 11 of 18 cases, the best of all. A detailed re-evaluation showed the disagreements were not random. Most involved florbetapir scans, and cortical atrophy was flagged as a contributing factor in four, while other cases reflected reduced image quality, subtle or borderline tracer uptake, or atypical regional distribution. Disagreement, in short, clustered in technically and biologically challenging images rather than revealing a flaw in either human or machine reading.

The authors are careful to frame these subgroup findings as descriptive rather than definitive, given the small sample, and they acknowledge that the higher hazard ratio for assisted interpretation at the cohort level may seem at odds with its weaker performance on the hardest cases. Their resolution is nuanced: quantitative tools appear to sharpen prognostic stratification across a whole clinic population, likely by nudging borderline calls toward the correct side, but they offer little incremental value for expert readers confronting genuinely ambiguous scans. This mirrors prior multicentre work showing that disagreement in amyloid PET concentrates in scans with low or borderline amyloid burden, and earlier trials demonstrating that trained visual readers detect beta-amyloid pathology with high sensitivity and specificity, including autopsy-validated correlations between florbetapir signal and plaque density.

The implications reach well beyond the reading room. As anti-amyloid immunotherapies move into routine practice, amyloid PET is no longer an academic exercise; it gates treatment decisions and, increasingly, therapeutic monitoring, with a growing trend toward reporting Centiloid values that harmonize quantification across tracers. The product characteristics of all three radiotracers already recommend quantitative analysis as a complement to visual assessment, and this study supports that framework while adding a caution: software overlays and machine-learning classifiers now emerging for automated amyloid classification should be treated as decision-support tools, not replacements for expert oversight, especially when atrophy, motion, or borderline uptake complicate the picture. The study’s limitations, including the absence of autopsy confirmation, the use of tracer-specific thresholds rather than Centiloid conversion, and a single institutional reconstruction protocol, leave room for refinement. But the core message is clear and reassuring: in the real world, a positive amyloid PET scan in a patient with amnestic mild cognitive impairment is a robust, method-independent warning of Alzheimer’s disease on the horizon, and the trained human eye remains an indispensable part of reading it.

Subject of Research: Prognostic value of amyloid PET interpretation methods for predicting Alzheimer's disease conversion in amnestic mild cognitive impairment

Article Title: Prognostic value of Amyloid PET in Real-World Amnestic MCI: Influence of reader experience and interpretation method

Article References: Echeveste, B., Guillén, E. F., Prieto, E., Hirschmüller, K. C. E., Riverol, M., & Arbizu, J. (2026). Prognostic value of Amyloid PET in Real-World Amnestic MCI: Influence of reader experience and interpretation method. European Journal of Nuclear Medicine and Molecular Imaging. https://doi.org/10.1007/s00259-026-08159-9

Image Credits: AI Generated

DOI: 10.1007/s00259-026-08159-9

Keywords: Alzheimer's disease, amnestic mild cognitive impairment, amyloid PET, visual interpretation, SUVR, z-score maps, inter-observer agreement, reader expertise, prognosis, nuclear medicine, biomarkers, dementia

The research team, led by Luke Borst of Antech Diagnostics and collaborators at Mars Petcare Science and Diagnostics, WALTHAM Petcare Science Institute, and Mars digital technology groups, developed and validated a deep learning system called OncoPetNet. The system is an ensemble of three fully convolutional neural networks built on encoder-decoder architectures with skip connections, a design descended from the U-Net family of image segmentation models. Rather than detecting mitotic figures directly as objects, the model frames the problem as segmentation: it predicts a pixel mask marking candidate dividing cells, and deterministic post-processing steps then convert those masks into coordinates. The encoders draw on EfficientNet-b5, EfficientNet-b3, and SE-ResNext backbones pre-trained on ImageNet, allowing the network to leverage general visual features learned from millions of natural images before fine-tuning on tumor histology.

Deploying such a model in a high-volume diagnostic laboratory required more than a clever network. The team built a production pipeline that runs asynchronously on every whole slide image uploaded to the lab. Each slide first passes through a biopsy slide classifier that decides whether mitotic counting is even appropriate. A non-model tissue detection algorithm then maps the foreground tissue and generates coordinates for the mitotic figure detector, which analyzes overlapping tiles across the entire tumor area. Morphological filters sharpen the predicted masks before a k-d tree based algorithm searches all possible placements of a 2.37 square millimeter square and selects the one containing the most mitotic figures. The final output, delivered to pathologists as annotations on the digital slide, includes both the region of interest and circles around each candidate mitosis.

Validation rested on two ground truth datasets. The first comprised 72 soft tissue sarcomas from the lab’s routine caseload, each confirmed by three pathologists, with consensus expert opinion establishing which AI-flagged candidates were true mitoses. On this dataset, OncoPetNet achieved an F1 score of 0.86, with recall, or sensitivity, reaching 0.97. The second benchmark was the MIDOG++ dataset, a publicly available multi-domain collection built specifically to test how well mitosis-detection models generalize across scanners, stains, laboratories, and tumor types. On 100 canine soft tissue sarcoma images from that set, the model scored an F1 of 0.70, with precision of 0.81 but recall dropping to 0.62. The authors attribute the difference partly to domain shift, since their model was trained on their own scanners and staining protocols, and partly to the low mitotic density of many MIDOG++ images, where a single missed mitosis can crater the score.

With the model validated, the team ran two evaluation studies. The first mimicked routine diagnostics: a single pathologist graded all 72 tumors three ways, manually, with AI assistance, and using AI alone. The results were striking. The AI-only condition averaged 19.9 mitotic figures per case, the AI-assisted condition 14, and the manual condition just 6.26. In other words, AI assistance allowed the pathologist to find about 2.24 times as many mitoses as unaided observation. The pathologist’s role shifted from exhaustive searching to verification: rejecting false positives, which the model produced more often than false negatives, and adding back any true mitoses the network missed. Human oversight measurably moderated the AI’s counts, which is precisely the behavior a safety-conscious deployment would want.

Those increased counts rippled through the grading scheme. Mitotic counts are binned into an ordinal MC score from 1 to 3, which is then summed with scores for tumor necrosis and degree of differentiation to produce the final grade. In the first study, AI-assisted scores migrated upward by one level in 29 percent of cases and by two levels in 6 percent compared with manual counts. Tumor grades followed: 18 percent of tumors moved from grade 1 to grade 2, and 7 percent from grade 2 to grade 3. No tumor migrated two full grades, and no tumor moved downward. The second study, in which nine pathologists organized into three cohorts each reviewed 24 cases both manually and with AI assistance, replicated these findings almost exactly. Across 216 paired observations, AI assistance produced an average of 11.6 mitoses versus 5.31 manually, again a roughly 2.2-fold increase, with grade migration of 15.7 percent from grade 1 to 2 and 5.6 percent from grade 2 to 3.

Efficiency gains were equally consistent. Manual counts took an average of 62.8 seconds per case, while AI-assisted counts took 33.7 seconds, nearly a 1.86-fold reduction. In a busy diagnostic laboratory processing thousands of submissions, halving the time spent on one of the most tedious steps in tumor grading represents a substantial operational improvement. Just as importantly, the AI’s pre-screening of every slide and field relieves pathologists of the cognitive burden of an unhurried, exhaustive search, a practice the authors note is technically required for accuracy but impractical under real-world workloads.

The most consequential findings concern agreement between pathologists. Using intraclass correlation coefficients, the team compared how consistently the nine pathologists rated the same tumors under manual versus AI-assisted conditions. For raw mitotic counts, AI assistance significantly improved agreement in two of the three pathologist trios and overall, lifting the overall coefficient from 0.94 to 0.98. Agreement on MC scores also improved significantly overall. Agreement on which slide to select, measured with Fleiss’ Kappa, rose in all three groups and reached statistical significance in two, reflecting the fact that AI effectively removes the guesswork about where to count. Agreement on final tumor grade showed a consistent upward trend, including a substantial jump in one group from 0.48 to 0.72, but these gains did not reach statistical significance. The authors attribute this dilution to the two remaining subjective parameters, percent necrosis and degree of differentiation, which AI does not yet assess and which continue to drive inter-pathologist variability.

The study’s authors are careful about what these results do and do not prove. The clinical significance of the higher, more sensitive counts remains unexplored: it is possible that current MC score thresholds, calibrated on imperfect manual counts, should be adjusted upward now that detection is better, or that AI-assisted grades will ultimately predict biological behavior more accurately. Testing that will require studies correlating AI-assisted grades with patient outcomes, ideally including clinical follow-up and necropsy data. Extending the approach to quantify necrosis and differentiation objectively would complete the standardization picture. There are also inherent limits to ground truth itself, since even expert consensus on what constitutes a mitotic figure carries some subjectivity, and immunohistochemical markers like phosphorylated histone H3 introduce their own interpretive complications.

What the study does establish is a working template for human-AI collaboration in diagnostic pathology. The machine does what machines do best, exhaustively scanning every slide and field without fatigue, while the human does what humans do best, exercising judgment to reject imposters such as apoptotic bodies, pyknotic nuclei, and inflammatory cells that mimic mitoses. The result is a count that is more sensitive, faster, and more reproducible than either party achieves alone. For veterinary oncology, where soft tissue sarcomas are among the most common canine tumors and grading decisions directly guide surgery and follow-up, that combination could mark the beginning of a genuine standardization of tumor grading, one algorithm-assisted count at a time.

Subject of Research: AI-assisted mitotic counting for standardizing histologic grading of canine soft tissue sarcomas

Article Title: Artificial intelligence-assisted mitotic counts improve efficiency, accuracy, and inter-pathologist agreement: a method toward canine soft tissue sarcoma grade standardization

Article References: Borst, L., Bacmeister, C., Yau, W., Aceino, A., Ranck, R., Labelle, P., Ehrensing, G., Oliveira, F., Myers, C., Willcox, J. L., Ellerby, Z., Haydock, R., Taylor, K., Parkinson, M., Fitzke, M., & Gerdin, J. (2025). Artificial intelligence-assisted mitotic counts improve efficiency, accuracy, and inter-pathologist agreement: a method toward canine soft tissue sarcoma grade standardization. Veterinary Oncology, 2(1), Article 19. https://doi.org/10.1186/s44356-025-00034-3

Image Credits: AI Generated

DOI: 10.1186/s44356-025-00034-3

Keywords: artificial intelligence, deep learning, veterinary oncology, soft tissue sarcoma, mitotic count, tumor grading, digital pathology, inter-observer agreement, canine cancer, convolutional neural networks, histopathology, MIDOG++