The Foundation Layer of AI Pathology: How Standardised Slide Preparation Sets the Ceiling for Artificial Intelligence

Industry commentary · October 2026

Over the past two years, pathology foundation models, vision-language models and pathology agents have taken turns on the stage of Nature and Nature Medicine, with ever more parameters and ever more tasks. Yet a plain engineering truism keeps being confirmed: the ceiling of an algorithm is set by the quality of its data — and that quality is already decided in the slide-preparation room.

For liquid-based cytology (LBC), this is especially true. Drawing on published research, this article reviews five dimensions — unmet need, quality control, traceability, data supply and clinical evidence — to set out why fully automated liquid-based thin-layer slide preparation is the most upstream and most easily under-estimated piece of infrastructure in AI pathology.

1. The overlooked starting point: slide quality was once the biggest gap in cervical screening

Conventional Pap smears opened the era of cervical cancer screening, but their technical ceiling was equally clear. A review published in Reviews in Obstetrics & Gynecology noted that conventional Pap smears show a wide reported range of sensitivity for high-grade lesions and a substantial false-negative rate, and that a considerable proportion of false negatives can be attributed to limitations in sampling or slide preparation — in other words, many missed diagnoses are not missed by the reader but by the slide itself: overlapping cells, obscuring mucus and red blood cells, and most of the collected cells being discarded with the sampling brush, so that only a fraction of the specimen reaches the slide^[1].

Liquid-based thin-layer technology was developed to address exactly this: cells enter a preservative solution, debris is removed by centrifugation or filtration, and the cells are then evenly deposited as a thin monolayer. The effect is measurable in controlled studies:

• A controlled study published in the Pakistan Journal of Medical Sciences reported a higher specimen adequacy rate with the liquid-based method than with conventional smears, and an increase in detection sensitivity for low-grade squamous intraepithelial lesions (LSIL) from 71.8% to 87.2%, and for high-grade lesions (HSIL) from 61.9% to 76.2%^[2].
• Laboratory survey data from the College of American Pathologists (CAP) showed higher median LSIL detection rates for liquid-based technologies (ThinPrep, SurePath) than for conventional smears^[1].

Liquid-based technology thus became the mainstream method in cervical cancer screening and in several other exfoliative cytology applications. But that is only half of the story.

2. The semi-automated compromise: people remain the largest variable in the system

Early liquid-based instruments were largely semi-automated: manual mixing, manual transfer, batch processing. Small differences between operators — in mixing force, settling time or stain volume — produce batch-to-batch and even slide-to-slide variation in cell density, distribution uniformity and staining intensity.

For human reading, such variation can often be compensated by experience. Its effect on AI is more direct. A study by a team at Shaanxi Provincial People's Hospital, published in the Journal of Modern Laboratory Medicine (2023), provided a telling comparison: the same AI-assisted system performed differently on membrane-based (TCT) slides and on sedimentation-based (LCT) slides^[3]. The study indicates that the preparation method alone can change the performance of the same AI model.

This is why the field moved early. In 2021, the relevant groups of the Chinese Society of Pathology published a Chinese expert consensus on digital pathological image acquisition and image quality control for cervical liquid-based cytology in the Chinese Journal of Pathology, raising image acquisition and pre-analytical quality control to a matter of standardisation^[4]. In 2024, the Chinese Journal of Diagnostic Pathology published experience with fully automated liquid-based thin-layer processors, noting their value in reducing human variation and stabilising slide quality^[5]. The consensus is clear: before talking about AI, get slide preparation under control.

3. Value one: moving quality control from retrospective sampling to a closed loop

A fully automated liquid-based thin-layer preparation line integrates sample identification, mixing, enrichment, sedimentation, independent staining, automated coverslipping and preliminary quality control into one system. In essence, it embeds quality control into every physical and software step: barcode identification at the input reduces the risk of sample mix-ups and omissions; programmed centrifugation and robotic pipetting reduce variation in manual force; controlled settling-chamber volume, temperature and time help deliver a uniform monolayer; independent staining chambers with single-use reagent help reduce cross-contamination and batch-to-batch variation; and sensors with visual positioning monitor the process in real time and can flag slides that fall outside limits.

This is no longer a concept but a documented practice. A study by the pathology department of a tertiary hospital in Nanning, published in the Chinese Journal of Diagnostic Pathology (2025), reported that after building and implementing a full-process quality-control system for liquid-based cytology, the slide excellence rate rose from 96.76% to 99.39% and the re-preparation rate fell from 1.07% to 0.12%^[6].

For AI, the significance of such figures is this: when the great majority of slides "look equally good", a model does not have to spend capacity learning to tolerate noise, and can devote more of its capacity to recognising genuine lesion features.

4. Value two: end-to-end traceability turns each slide into an auditable data asset

For AI pathology to enter hospitals, pass regulatory scrutiny and carry responsibility, one question must be answerable: where did this image come from? Through "one vial, one code, one code to the end", a fully automated line writes the sample-vial barcode, the laser-etched slide code, HIS/LIS case information, mixing time, centrifugation speed, reagent lot, coverslipping temperature, operator and consumable batch number into a structured log — turning each slide into a traceable, auditable data asset.

The academic value of this capability is now being formally recognised. A framework study published in the Journal of Pathology Informatics (2026) noted that computational pathology datasets often conflate biological units (patients, specimens), laboratory preparations (blocks, slides) and derived digital data (whole-slide images, tiles), and recommended following the DICOM hierarchical information model to report counts explicitly at each level from patient and specimen through block and slide to whole-slide image and tile^[7]. Fully automated preparation standardises how data is generated and recorded at the source, making such slide-level transparency achievable — a prerequisite for aligning multi-centre validation, real-world studies and federated learning.

5. Value three: supplying standardised input for foundation models

Pathology AI has entered the foundation-model era, and foundation models grow by eating data:

• PathOrchestra, published in npj Digital Medicine (2025), was pre-trained by self-supervision on 287,424 whole-slide images from 3 centres and 21 tissue types, and validated on 112 clinical-grade tasks^[8].
• TITAN, from the Mahmood laboratory at Harvard Medical School and published in Nature Medicine (2025), was pre-trained on 335,645 whole-slide images paired with more than 180,000 pathology reports and over 420,000 synthetic descriptions, enabling retrieval and report generation without fine-tuning^[9].

What these models share is a prerequisite: large volumes of high-quality, standardised image input. Uneven cell distribution, background debris and staining fluctuation all enter the feature space as noise and weaken a model's learning of genuine pathological features.

More direct evidence comes from a systematic experiment published in the Journal of Pathology: Clinical Research (2026): using soft-tissue tumours as a test case, the authors assessed the robustness of seven AI models — including UNI-v2, Virchow and TITAN — to staining variation and to differences between scanner brands. Foundation models showed a degree of robustness, but model performance improved markedly when a small number of slides with staining variation were included in the training loop^[10]. The converse of that finding is precisely the point: the more standardised the data source, the more reliable deployment becomes in the real world — and homogenising preparation upstream is among the most economical and most fundamental ways to reduce this class of non-biological variation.

6. Evidence from China: AI cytology works — provided it is fed well

In cervical cytology, the fastest-moving subspecialty for AI adoption, Chinese studies have accumulated substantial evidence:

• A study by a team at Shengjing Hospital of China Medical University on 29,354 cervical cytology smears (Journal of China Medical University, 2023) reported that, with review results as the final diagnosis, AI diagnostic sensitivity reached 98.59% and negative predictive value 99.66%; the AI miss rate (1.41%) was lower than that of manual reading (2.69%), with no missed LSIL-or-above cases; and AI took less time per slide than a pathologist^[11].
• A study by a team at the Third Affiliated Hospital of Zhengzhou University, published in Modern Pathology (2024), reported that with an AI-assisted system (AICyte), pathologists' average reading time fell substantially, sensitivity for CIN2+ lesions matched the original diagnosis (97.89%) and specificity improved^[12].
• An earlier multi-centre cohort study (Cancer Medicine, 2020) validated an AI-assisted cytology diagnostic system in a nationwide screening population of approximately 0.7 million women^[13]. The AICCS system developed at Sun Yat-sen Memorial Hospital of Sun Yat-sen University, published in Nature Communications (2024), confirmed in multi-centre and prospective datasets that AI assistance improved cytopathologists' AUC, sensitivity and specificity^[14]. That study also noted that more than 350 million women aged 30–65 in China are expected to need regular cervical cytology screening, far exceeding current capacity^[14].

Note, however, the condition implicit in all of these studies: they rest on liquid-based thin-layer slides of adequate quality. As the comparison study above^[3] indicates, once preparation quality fluctuates, AI performance drops. Reproducing study-level performance in every hospital and every batch is not a matter of luck but of standardised preparation infrastructure.

7. Conclusion: the starting point of intelligence is written on the slide

Turning pathology pre-analytical work from a craft that depends on individual experience into an engineering system that is quantifiable, reproducible and traceable — that is the real significance of a fully automated liquid-based thin-layer preparation line:

• Quality control: making every slide "good enough". The study cited above shows an excellence rate of 99.39% and a re-preparation rate of 0.12% are attainable^[6]; how far other institutions reach depends on their own conditions.
• Traceability: making every slide "trustworthy". Complete hierarchical records from patient to tile are precisely the data standard the international community is calling for^[7].
• Standardisation: enabling those slides to genuinely feed foundation models and diagnostic agents, so that AI moves from laboratory demonstration to clinical infrastructure.

When a pathology department truly achieves "sample in, report out, with AI quality control and assistance throughout", it becomes clear that the starting point of all intelligence is written on that standardised thin-layer slide. The algorithm decides how fast AI can run; slide preparation decides how far AI can go.

References

1. The Impact of Liquid-Based Cytology in Decreasing the Incidence of Cervical Cancer. Reviews in Obstetrics & Gynecology.
2. Comparison of PAP smear and liquid based cytology as a screening method for cervical carcinoma. Pakistan Journal of Medical Sciences. Journal page
3. Chen J, Li WS, Zhang W. Application value of an AI-assisted system in cervical liquid-based cytology analysis. Journal of Modern Laboratory Medicine, 2023, 38(5): 155-159. (In Chinese.)
4. Digital Pathology and AI Pathology Group, Chinese Pathologists Association, et al. Chinese expert consensus on digital pathological image acquisition and image quality control of cervical liquid-based cytology. Chinese Journal of Pathology, 2021, 50(4): 319-322. (In Chinese.) Journal page
5. Su HD, Gao LL, Li Q. Experience with a fully automated liquid-based thin-layer cell processor. Chinese Journal of Diagnostic Pathology, 2024, 31(7): 680-681+684. (In Chinese.)
6. Li BQ, Mo YQ, Xie L. Value of a full-process quality control system for liquid-based cytology. Chinese Journal of Diagnostic Pathology, 2025(6). (In Chinese.)
7. Dy A, et al. A straightforward framework to harmonize computational pathology. Journal of Pathology Informatics, 2026. Article
8. Yan F, Wu J, Li J, et al. PathOrchestra: a comprehensive foundation model for computational pathology with over 100 diverse clinical-grade tasks. npj Digital Medicine, 2025, 8: 695. Article
9. Ding T, Wagner SJ, Song AH, et al. A multimodal whole-slide foundation model for pathology (TITAN). Nature Medicine, 2025, 31: 3749-3761.
10. Chai B, Chen J, Cool P, et al. Impact of tissue staining and scanner variation on the performance of pathology foundation models: a study of sarcomas and their mimics. Journal of Pathology: Clinical Research, 2026, 12(2): e70080. DOI: 10.1002/2056-4538.70080
11. Zhang SW, Zhang WC, Zhang Z, et al. Application of artificial intelligence in preliminary screening of cervical lesions. Journal of China Medical University, 2023, 52(8): 694-698. (In Chinese.) Journal page
12. Assessment of Efficacy and Accuracy of Cervical Cytology Screening with Artificial Intelligence Assistive System (AICyte). Modern Pathology, 2024.
13. Bao H, Sun X, Zhang Y, et al. The artificial intelligence-assisted cytology diagnostic system in large-scale cervical cancer screening: a population-based cohort study of 0.7 million women. Cancer Medicine, 2020, 9(18): 6896-6906. PubMed
14. Wang J, Yu Y, Tan Y, et al. Artificial intelligence enables precision diagnosis of cervical cytology grades and cervical cancer (AICCS). Nature Communications, 2024, 15: 4369. Article