Investigation report: "A fully automatic AI system for tooth and alveolar bone segmentation from cone-beam CT images" (Cui et al., Nature Communications, 2022)

Academic fraud report · Geng Detector

Summary

Verdict: Questionable 🟡. The report identifies three issues in the paper published in Nature Communications (DOI: 10.1038/s41467-022-29637-2) concerning a deep learning system for CBCT tooth/bone segmentation. Finding 1 flags an internal inconsistency between the Methods text and Table 1 / Fig. 1b: the text attributes the total counts of 4,938 CBCT scans and 4,215 patients to the internal dataset alone, but the summed components (CQ + HZ + SH) yield only 4,531 scans and 3,811 patients; the totals are reached only when external data is included. Finding 2 notes a mismatch between the text ("147 and 169 min" for the two radiologists) and Table 3 (147 and 160 min). Finding 3 raises concerns about imprecise use of the term "resolution" in the image pre-processing section. No image manipulation or data fabrication is alleged; all issues appear consistent with sloppy writing and cross-checking.

Verdict

🟡 Questionable. The identified problems are consistent with textual errors and unclear methodological writing rather than systematic data fabrication. Publication of a formal correction or corrigendum is recommended.

Key findings

Internal vs. total dataset mismatch (Finding 1): Methods states that the internal dataset alone contains 4,938 CBCT scans of 4,215 patients from CQ-, HZ-, and SH-hospital. However, summing the per-hospital numbers in Table 1 / Fig. 1b gives only 4,531 CBCT scans (1,532 + 1,798 + 1,201) and 3,811 patients (924 + 1,689 + 1,198). The totals 4,938 / 4,215 are only reached when the external dataset (407 scans / 404 patients) is added.
Radiologist annotation time mismatch (Finding 2): Text reports "147 and 169 min" (average) for the two expert radiologists, but Table 3 reports "147" and "160" min. A 9-minute discrepancy for the second radiologist is not explained.
Imprecise terminology in image pre-processing (Finding 3): The Methods describe resampling decisions using "higher than 0.4 mm" as the criterion, conflating physical voxel spacing with "resolution." In CBCT imaging, a larger spacing (e.g., 0.5 mm) corresponds to lower, not higher, resolution, indicating sloppy description rather than a methodological defect.

Evidence highlights

Per-hospital CBCT scan counts from Table 1 / Fig. 1b: 1,532 (CQ) + 1,798 (HZ) + 1,201 (SH) = 4,531 (internal), vs. stated 4,938 in text.
Per-hospital patient counts: 924 + 1,689 + 1,198 = 3,811 (internal), vs. stated 4,215 in text.
External dataset adds 407 scans / 404 patients, matching the stated totals (4,531 + 407 = 4,938; 3,811 + 404 = 4,215).
Table 3, Time (min): 147 and 160; text states 147 and 169.
DOI: 10.1038/s41467-022-29637-2.

Notes

All discrepancies are confined to textual/descriptive content; the underlying algorithmic results and figure visuals are not implicated.
The most likely explanation is a writing/editing error where the authors omitted that the totals (4,938 scans, 4,215 patients) include the external dataset, and a typo in the second radiologist's annotation time.
These issues warrant a corrigendum but do not, on current evidence, meet the threshold for a misconduct referral.
Confidence in the arithmetic mismatch (Finding 1) is high; confidence in the "169 vs. 160" mismatch (Finding 2) is high given direct table evidence; confidence in the terminology critique (Finding 3) is moderate, as it reflects imprecise exposition rather than a technical error.