Integrity Review Report: "Enhancing representation in radiography-reports foundation model: a granular alignment algorithm using masked contrastive learning" (Huang et al., Nature Communications, 2024)

Academic fraud report · Geng Detector

Summary

This report raises four substantive concerns against the above Nature Communications paper, with an overall verdict of highly suspicious. Finding 1 documents an institutional name error: Harvard Medical School is rendered as "Harvard Medical University," an entity that does not exist. Finding 2 identifies an arithmetic impossibility in the RSNA Pneumonia Detection dataset description, where 25,184 + 1,500 + 3,000 = 29,684 exceeds the official ~26,684 images by 3,000. Finding 3 highlights implausibly large performance jumps on COVID Rural segmentation (MaCo at 75.1% Dice vs. the previous best of 44.0% MedKLIP at 100% labels, a >70% relative gain) without adequate ablations. Finding 4 flags implausible compute claims: 3.5 hours on four A100 GPUs to pretrain on 377,000 image-report pairs, which is inconsistent with known training times for ViT/BERT-scale multimodal models. No image manipulation was assessed due to lack of figures. The findings are based on text, arithmetic and logical analysis; final determination of misconduct requires institutional investigation.

Verdict

Highly suspicious. Multiple basic factual, arithmetic, and computational-plausibility issues were identified, none of which appear individually minor; together they suggest either fabrication, extreme carelessness, or outsourced writing without subject-matter oversight.

Key findings

Non-existent institutional affiliation (severity: critical). The author Hong-Yu Zhou is listed under "Department of Biomedical Informatics, Harvard Medical University, Boston, MA, USA." No such institution exists; Harvard's medical school is Harvard Medical School (HMS).
Dataset arithmetic inconsistent with public record (severity: critical). RSNA Pneumonia Detection is reported as split into 25,184 training + 1,500 validation + 3,000 test images, totaling 29,684, whereas the official dataset contains approximately 26,684 images — a discrepancy of 3,000 images.
Implausible performance gain (severity: high). On COVID Rural segmentation, MaCo reports 75.1% Dice versus the prior best of 44.0% (MedKLIP) at 100% labels, a >30 absolute-point / >70% relative improvement over eight SOTA baselines, discussed in a single sentence without ablations or failure-case analysis.
Implausible training cost claim (severity: high). Pre-training on >377,000 image-report pairs of MIMIC-CXR V2 with a ViT-B/16 image encoder, BERT text encoder, masked autoencoding and contrastive objectives is stated to take ~3.5 hours on four NVIDIA A100 GPUs, which is inconsistent with reported training durations for comparable multimodal foundation models.

Evidence highlights

Affiliation text: "Department of Biomedical Informatics, Harvard Medical University, Boston, MA, USA." (Page 1 footnote). Harvard Medical University is not a recognized institution; Harvard Medical School (HMS) is the correct name.
Dataset numbers (Page 9, Methods — Datasets, RSNA Pneumonia Detection paragraph): "training set of 25,184 images, a validation set of 1500 images, and a test set of 3,000 images." Sum = 29,684 vs. official ~26,684.
Table 3 (Page 4, fine-tuning segmentation): COVID Rural 100% labels — MaCo Dice 75.1 vs. MedKLIP 44.0; previous SOTA range cited across eight methods.
Implementation details (Page 8): "The pre-training of MaCo was completed in approximately 3.5 hours using four NVIDIA A100 GPUs" on MIMIC-CXR V2 with >377,000 image-report pairs.
DOI: 10.1038/s41467-024-51749-0

Notes

Scope: this review is limited to text, arithmetic and logical plausibility; no image-based duplicate or splicing analysis was possible because figures were not provided for pixel inspection.
The institutional-name error and the RSNA count error are independently verifiable and, if confirmed against the published version, would constitute basic factual mistakes of a kind not expected from authors with genuine primary access to the data and institutions claimed.
The two high-severity findings (segmentation jump and training-time claim) are indicators rather than proof; they require re-running experiments with disclosed code, seeds and logs.
Authors and editors have not yet been contacted as part of this report; recommended follow-ups include requesting raw splits and training logs, posting on PubPeer, notifying the Nature Communications editorial office, and verifying Zhou's affiliation with Harvard Medical School's Office of Academic Affairs.