Academic Integrity Review: Deep Learning-Based Autofocus Algorithm for Droplet Digital PCR (Chinese Journal of Lasers, DOI 10.3788/CJL251660)

Academic fraud report · Geng Detector

Summary

This report examines "基于深度学习的微滴数字PCR自动对焦算法研究" (Deep learning-based autofocus for droplet digital PCR) by Qu Yanheng et al., Xi'an Jiaotong University, Chinese Journal of Lasers, DOI 10.3788/CJL251660. Verdict: highly suspicious but not convicted. Four downgraded concerns remain: (1) validation accuracy curves for learning rates 0.0005 and 0.0001 nearly coincide after convergence (Fig. 7(b)); (2) focusing-time curves in Fig. 9(d)-(f) show perfect V/U symmetry across three methods despite claimed single-shot times of 34-64 s, 35-70 s and 5-8/5-11 steps; (3) clarity-evaluation curves appear overly smooth, though the paper notes the metric flattens beyond |Δfocus|>900; (4) dataset arithmetic (549×89=48861, minimum ~97722 after 2× augmentation) cannot reconcile with the stated 65340 training samples (split 45738/13068/6534). Automated Benford/last-digit tests and copy-move/nnoise flags were cleared as inapplicable or benign. A Bayesian synthesis reports 92.8% posterior, but all individual findings lack numeric/textual confirmation; original training logs, timing records, and dataset construction details are required before any conclusion.

Verdict

Highly suspicious (🟠) — recommend investigation; no finding individually sufficient for a determination of misconduct. All flagged findings are downgraded pending original data. The automated Bayesian synthesis (posterior 92.8%, BF = 246.55, 95% CI [66.9%, 98.5%]) reflects the *cumulative* pattern of visual/anomaly suspicions, not confirmed fraud. Statistical and image-forensics alerts (Benford, last-digit, copy-move, PRNU) were reviewed and cleared as inapplicable or benign.

Paper: 基于深度学习的微滴数字PCR自动对焦算法研究 — Qu Yanheng, Zhao Shuhao, Ren Zhenliang, Peng Niancai, Zhang Zhenxi, Hu Fei (Xi'an Jiaotong University). Chinese Journal of Lasers, Vol. 53, No. 15, 2026 (received 2025-12-22). DOI: 10.3788/CJL251660

Key findings

Fig. 7(b): Validation accuracy curves for initial learning rates 0.0005 and 0.0001 nearly coincide after convergence, despite the paper claiming independent training runs (l = 0.01, 0.001, 0.0005, 0.0001, 0.00001). Visual concern only; convergence to similar accuracy is plausible in deep learning. No numeric evidence; requires training logs.
Fig. 9(d)-(f): Focusing-time curves for the coarse-fine method, conventional method, and simulated annealing all show perfect V/U-shaped symmetry versus initial defocus distance, inconsistent with typical asymmetric hardware characteristics. Paper claims measured single-shot times of 34–64 s / 35–70 s and 5–8 / 5–11 steps. No smoothing/fitting described; requires raw timing records.
Fig. 2(a), Fig. 8(a): Clarity-evaluation curves overly smooth with flat baselines (Focus < 1000 and > 2500); however, the paper explicitly states the metric change is negligible when |focus offset| > 900, providing a benign explanation for the flat baseline. Peak-region smoothing still unverified.
Dataset arithmetic (Sections 2.1 / 3.1.2 / 3.2.3): Claimed 549 images (61×9), each cropped to 89 sub-images, augmented to 65340 samples, split 7:2:1 (45738/13068/6534 = 65340, internally consistent). But 549×89 = 48861, ×2 minimum = 97722 ≥ 65340; using only |defocus| < 900 images (≈333) yields 59274, leaving a 6066 gap requiring ~20.5% of images to be boundary 4×-augmented — unreported. Cannot be reconciled from the text; benign undisclosed filtering possible.
Cleared alerts: Benford (χ² = 121.5/201.0/564.2/507.8) and last-digit tests (χ² = 96.7; 13.2σ) invalid for engineered/controlled parameter data. Copy-move/nnoise flags (~10 pages, e.g., offsets (32,0), (40,0), (0,32)) consistent with multi-panel composite figures confirmed by captions; PRNU NCC near zero (weak). Retained only as manual-review leads.

Evidence highlights

All plotted data points passed programmatic verification against text claims where stated (e.g., 34–64 s, 5–8 steps, 65340 samples).
Bayesian synthesis: posterior 92.8%, 95% CI [66.9%, 98.5%], BF = 246.55; top contributors are findings 6, 3, 4, 1, 2 (each logLR ≈ 1.08–1.19, ~19–22%). Assumption of conditional independence may overstate joint strength; calibration n = 11 (LR version geng_lr_v3).
No finding has numeric-level (raw-data) confirmation; all four open findings are visual or chain-of-custody anomalies with possible innocent explanations.

Notes

Recommended actions: contact authors for original data (focusing-time records, curve computation values, training logs/accuracy history, dataset construction details, original high-resolution images); escalate to journal/institution if unavailable.
Uncertainty: findings 1, 2, 4 are visual/arithmetic suspicions only; finding 3 has explicit textual support for benign explanation; finding 6 flags require per-panel pixel-level review.
AI-assisted report for academic discussion only; final misconduct determination requires official institutional investigation.