Screening efficiency over experience: Rapid target detection in low-power field as a modifiable cognitive biomarker for diagnostic accuracy in digital cytology
Eye-tracking study of 100 cytotechnologists (1–40 years experience) and 28 students evaluating 30 digital cytology images. Demonstrates the 'experience paradox': professional experience showed no sign
Abe N, Nishimura Y, Yamashita K, Kawamorita T, Takatori Y, Murakumo Y, Furuta R. Screening efficiency over experience: Rapid target detection in low‐power field as a modifiable cognitive biomarker for diagnostic accuracy in digital cytology. Cancer Cytopathology (2026). Published online: 7 July 2026. DOI: 10.1002/cncy.70132.
Wiley Online Library: doi.org/10.1002/cncy.70132
Local PDF:
K:\DownloadsK\Cancer Cytopathology - 2026 - Abe - Screening efficiency over experience Rapid target detection in low‐power field as a.pdfInstitution: Kitasato University (School of Allied Health Sciences, Graduate School of Medical Sciences, and Hospital), Kanagawa, Japan.
Executive Summary
In classical light microscopy, screening expertise has traditionally been equated with accumulated years of clinical on-the-job experience. However, the adoption of digital whole-slide imaging (WSI) and artificial intelligence (AI) is fundamentally altering the human observer's role: instead of performing exhaustive manual tile-by-tile sweeps, pathologists and cytotechnologists (CTs) increasingly act as rapid verifiers of candidate regions of interest (ROIs) proposed by algorithms.
This landmark eye-tracking study evaluates gaze dynamics across two distinct cohorts:
Phase 1 (Cross-Sectional Expertise Study, $n = 100$ CTs): Evaluated 100 practicing cytotechnologists with professional experience spanning 1 to 40 years.
Phase 2 (Longitudinal Educational Intervention, $n = 28$ Trainees): Tracked 28 cytotechnology students before and after a standard 3-month practical training program.
The results establish two fundamental principles:
The Experience Paradox: Years of professional experience showed no significant correlation with diagnostic accuracy ($r = 0.189, p = .0596$). Seniority primarily correlated with time spent reading clinical requisition notes ($r = 0.228, p = .022$), reflecting cautious habits rather than superior discriminatory vision.
LPF Pop-Out Efficiency as Independent Predictor: In multivariate nominal logistic regression, shorter total fixation duration on the low-power field (LPF) main diagnostic target was the sole independent predictor of high diagnostic accuracy ($p = .0454$). High performers rely on global parafoveal processing ("pop-out" target detection) rather than slow, exhaustive focal scrutiny.
Trainable Selective Neglect: Trainees in Phase 2 acquired this expert visual pattern within 3 months, dramatically accelerating time to first target fixation ($r = 0.86, p < .0001$) while increasing the time to first fixation on normal background cells ($r = 0.71, p < .0001$). They learned what not to look at ("selective neglect" / information reduction).
Study Design & Experimental Methods
Areas of Interest (AOIs)
Using Tobii Pro Lab software, gaze positions were recorded across 6 calibrated AOIs:
Sample Information: Specimen source, patient age, and gender on the digital screen.
Background: Slide areas devoid of diagnostic cells or cellular clusters.
LPF Normal Cells: Benign/normal cell groupings visible at low magnification.
LPF Main Object: The primary diagnostic lesion/atypical cluster in the low-power view.
HPF Cytoplasm: High-magnification cytoplasmic texture and boundaries.
HPF Main Object: High-magnification nuclear chromatin, nucleoli, and nuclear membranes.
Detailed Findings
1. The "Experience Paradox"
Overall diagnostic accuracy across 100 CTs averaged 82.1% (range: 63.3%–96.7%; 95% CI: 80.6%–83.7%).
Years of experience did not correlate with diagnostic accuracy ($r = 0.189, p = .0596$).
While all participants with $<70%$ accuracy had $<10$ years experience, many novices achieved diagnostic accuracies comparable to or exceeding veterans with $20+$ to $40$ years of service.
What experience did predict: Senior CTs spent significantly more time reviewing the clinical history (first visit duration: 0.18 s vs. 0.14 s, $p = .026$; visit count: $r = 0.228, p = .022$). However, this compensatory caution did not translate into higher diagnostic yields.
2. Gaze Metrics Predicting High Diagnostic Accuracy
ROC analysis established an optimal diagnostic accuracy threshold of $>83%$ ($p = .0088$).
High-accuracy participants exhibited:
Significantly shorter total examination time ($p = .0076$).
Less background dwelling ($p = .016$).
Faster time to first fixation on diagnostic cytoplasm ($p = .039$).
In multivariate nominal logistic regression incorporating experience, sample information fixation, and target fixation:
Total fixation duration on the LPF main object was the only statistically significant independent predictor ($p = .0454$).
Practitioners who rapidly recognized abnormal patterns on low power and confirmed them swiftly with high power achieved superior accuracy.
Total Examination Time
Rapid (Mean ~14–25 s)
Prolonged (Up to 104 s)
$p = .0076$
Total Fixation on Background
Minimal
Recursive / Extended
$p = .016$
Time to HPF Target Fixation
Immediate
Delayed
$p = .039$
Search Pattern Structure
Linear, rhythmic (LPF $\to$ HPF)
Fragmented, cyclical loops
NVivo spatiotemporal
Multivariate Predictor
Short LPF Target Fixation
Long LPF Target Fixation
$p = .0454$
3. Spatiotemporal Gaze Dynamics (NVivo Visualizations)
Spatiotemporal timeline reconstructions showed two fundamentally divergent visual strategies:
High Performers (Linear Rhythmic Strategy): Completed slide evaluation in ~14 seconds. Immediate low-power target detection followed by a direct, unhesitating hop to high-power nuclear/cytoplasmic verification, concluding with final classification.
Low Performers (Fragmented Recursive Strategy): Required up to 104 seconds. Exhibited disjointed scanpaths, wandering over benign cellular clusters, repeatedly returning to requisition data, and re-checking irrelevant background areas.
4. Educational Modifiability & "Selective Neglect" (Phase 2)
In the longitudinal cohort of 28 cytotechnology students before and after a 3-month curriculum:
Rapid Target Acquisition: Time to first fixation on diagnostic cytoplasm decreased with a massive effect size ($r = 0.86, p < .0001$).
Acquisition of Selective Neglect: Time to first fixation on normal background cells significantly increased ($r = 0.71, p < .0001$). Trainees learned to actively disregard non-informative normal cells.
This demonstrates that visual expertise is not an inevitable byproduct of decades of unguided exposure, but a discrete, learnable cognitive mechanism—the strategic filtering of visual noise—that develops rapidly under structured training.
Implications for AI-Assisted Digital Pathology Workflows
The Shift from "Searcher" to "Verifier": Traditional cytopathology workflows required exhaustive visual rastering across thousands of fields of view to find isolated abnormal cells. In AI-augmented systems (e.g., automated cervical cytology screening algorithms flagging 20–50 candidate fields), the human operator's cognitive task is pure verification.
Inefficiency as an AI Liability: Practitioners who rely on slow, exhaustive scanning and frequent meta-data checking rather than rapid gestalt recognition will experience severe bottlenecks when tasked with verifying high-throughput AI candidate galleries.
Objective Gaze Metrics Over Time-Based Credentialing: Credentialing in pathology and cytotechnology has historically been defined by training duration and cumulative case counts. Eye-tracking gaze metrics (e.g., LPF time to target fixation, selective neglect ratio) provide objective, quantifiable cognitive biomarkers for diagnostic competency and AI readiness.
Countering Confirmation & Automation Biases: As shown in When Two Wrongs Don't Make a Right: Examining Confirmation Bias and the Role of Time Pressure During Human-AI Collaboration in Computational Pathology, time pressure drives practitioners into uncritical automation bias. Developing rapid, robust LPF verification skills protects against cognitive fatigue and unreflective acceptance of false-positive AI bounding boxes.
Limitations
Static Images vs. Dynamic WSI: Eye-tracking was conducted on 30 static digital image fields rather than continuous gigapixel panning/zooming. However, final diagnostic decisions fundamentally depend on static field interpretation.
Specimen Preparations: Focused on conventional Pap smears and respiratory cytology; liquid-based cytology (LBC) thin-layer preparations were not evaluated, though the authors note that global-focal cognitive mechanisms operate similarly across preparations.
Sample Size of Diagnostic Set: 30 cases per testing round.
Vault Relationships
Last updated