Abe 2026 — Screening Efficiency Over Experience in Digital Cytology

Eye-tracking study of 100 cytotechnologists and 28 students on 30 digital cytology cases — years of experience does not predict diagnostic accuracy (r = 0.189); instead, rapid low-power field target detection (‘pop-out’ mechanism) is the sole independent predictor (p = .045) and is rapidly learnable via selective noise neglect.

Author

Abe N, Nishimura Y, Yamashita K, Kawamorita T, Takatori Y, Murakumo Y, Furuta R

Doi

10.1002/cncy.70132

In digital cytology, diagnostic accuracy is governed by rapid low-power field target detection (“pop-out” efficiency) rather than accumulated years of professional experience (\(r = 0.189, p = .0596\)), and this visual efficiency develops rapidly through education via selective neglect of uninformative background cells — an empirical demonstration that human verification skill in digital workflows is a discrete, measurable, and modifiable cognitive biomarker.

Read status: full text read from local PDF. Captures the Phase 1 cross-sectional cohort of 100 cytotechnologists (1–40 years experience), Phase 2 longitudinal training evaluation of 28 students, Tobii Pro Lab gaze metrics, multivariate nominal logistic models, and NVivo spatiotemporal scanpaths.

The clinical, cognitive, and AI question

Traditional cytology instruction and certification rely on an apprenticeship paradigm where competency is assumed to accrue passively with years on the bench. However, the ongoing transition to digital whole-slide imaging and AI-assisted screening fundamentally shifts the human task: 1. From searcher to verifier: Instead of manually sweeping tens of thousands of fields, practitioners in AI-in-the-loop workflows review gallery crops or targeted regions of interest (ROIs) flagged by algorithms. 2. The cognitive bottleneck: Does decades of manual glass-slide experience translate into superior digital diagnostic accuracy? Or does it entrench exhaustive, slow scanning behaviors that become liabilities in high-throughput digital and AI verification environments? 3. Modifiability: Can the visual search efficiency characterizing top performers be formally taught and objectively quantified using gaze metrics?

The four fields

Dataset & Cohorts	Two-phase study at Kitasato University Hospital & Allied Health Sciences (Japan): • Phase 1 (Expertise analysis): \(n = 100\) certified cytotechnologists (CTs), professional experience 1–40 years (stratified by ROC at \(\le 5\) vs. \(>5\) years). Diagnosed 30 static digital cytology images (15 cervical, 15 respiratory). • Phase 2 (Educational intervention): \(n = 28\) cytotechnology students before and after a standard 3-month practical training program, evaluated on two matched sets of 30 cases (Set A pre, Set B post; 15 cervical, 15 respiratory).
Eye-tracking / AOIs	Tobii Pro Lab recording 6 calibrated areas of interest (AOIs): 1. Sample information (patient demographics, specimen site) 2. Background (cell-free zones) 3. LPF normal cells 4. LPF main object (diagnostic target on low magnification) 5. HPF cytoplasm 6. HPF main object (nuclear features on high magnification). Metrics: Time to first fixation, fixation count, fixation duration, visit count.
Statistical Analysis	Phase 1: Pearson correlation, ROC curves, Wilcoxon rank-sum, and multivariate nominal logistic regression to identify independent predictors of high accuracy (>83%). Spatiotemporal scanpaths mapped with NVivo. Phase 2: Paired Wilcoxon signed-rank tests and Cohen’s effect size (\(r = Z/\sqrt{N}\)).
Validation / Scope	Evaluated on conventional cervical and respiratory cytology smears. Ground truth established by expert consensus. Gaze tracker was purely passive (no live eye-tracking feedback was provided to trainees during learning).

Findings

1. The “Experience Paradox”

Overall CT accuracy: mean 82.1% (range 63.3%–96.7%).
Years of experience showed no significant linear correlation with diagnostic accuracy (\(r = 0.189, p = .0596\)).
Senior CTs (\(>5\) years) spent significantly more time reviewing requisition/sample metadata (first visit duration: 0.18 s vs 0.14 s, \(p = .026\); visit count: \(r = 0.228, p = .022\)). While reflecting ingrained caution, this habit had zero correlation with diagnostic success.

2. Low-Power Field (LPF) “Pop-Out” Detection as the Sole Predictor

Diagnostic accuracy correlated negatively with total examination time (\(r = -0.235, p = .018\)) and total fixation duration on the LPF main object (\(r = -0.211, p = .035\)). High performers made decisions faster and required less dwell time on targets.
In multivariate nominal logistic regression, total fixation duration on the LPF main object was the sole significant independent predictor of high accuracy (\(p = .0454\)).
This reflects the global-focal search model: expert readers detect diagnostic abnormalities pre-attentively via parafoveal vision (“pop-out” effect) at low magnification, requiring only swift, confirmatory high-power verification.

3. Spatiotemporal Scanpath Topologies

High performers (Linear/Rhythmic): Completed cases in ~14 seconds. Direct scanpath from slide onset to LPF target, instantaneous jump to HPF nuclear/cytoplasmic verification, and prompt diagnosis.
Low performers (Fragmented/Recursive): Took up to 104 seconds. Recursive wandering over normal cells and empty background, repeated regressions back to patient demographic headers, and prolonged decision hesitation.

4. Educational Modifiability and “Selective Neglect” (Phase 2)

After 3 months of training without eye-tracking interventions, students’ gaze dynamics shifted dramatically toward expert profiles: - Accelerated target acquisition: Time to first fixation on diagnostic HPF cytoplasm decreased markedly (\(p < .0001\), effect size \(r = 0.86\)). - Acquired selective neglect: Time to first fixation on normal LPF cells significantly increased (\(p < .0001\), effect size \(r = 0.71\)). Students learned information reduction—actively ignoring visually prominent but diagnostically irrelevant benign background noise.

Transferable takeaways for pathology, AI, and verification workflows

Exhaustive screening is not expertise: Unstructured years on the bench can entrench slow, exhaustive rastering habits. Expertise is defined by selective visual filtering and rapid target verification.
Direct relevance to AI-assisted cytology: As AI triage algorithms (e.g., automated Pap smear screeners) present pre-selected FOV galleries, the human role transitions entirely from exhaustive searcher to rapid verifier. LPF pop-out efficiency is the precise cognitive trait needed for high-throughput, accurate AI verification.
Objective competency metrics: Eye-tracking metrics (time to first target fixation, background fixation ratio) can replace arbitrary case-volume quotas with quantifiable cognitive biomarkers for trainee assessment and AI readiness.

How it connects

Cervical Cytology Screening — the high-volume screening domain where automated AI systems are already widely deployed and where rapid human verification is critical.
Cognitive Bias in AI-Assisted Diagnosis — expanding cognitive research in pathology beyond the Rosbach TCP studies to empirical gaze tracking, selective attention, and the cognitive mechanics of human-machine interaction.
Rosbach 2025 — Confirmation Bias and Time Pressure in AI-Assisted Tumour Cell Percentage Estimation — complementary evidence on human decision-making: while Rosbach showed how time pressure collapses confirmation bias into automation bias, Abe demonstrates how baseline visual efficiency dictates verification speed and accuracy.
Interobserver Agreement — illustrating that reader variability stems from divergent visual search architectures (rhythmic global-focal vs fragmented recursive) rather than just diagnostic threshold differences.