Integrity Review: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for CheX-ray Vision-Language Pre-training (DOI: 10.1145/3746027.3755336)

Academic fraud report · Geng Detector

Summary

This report assesses a 2025 ACM Multimedia paper (DOI: 10.1145/3746027.3755336) proposing a vision-language pre-training framework with a Pathology-Aware Reconstruction (PAR) module and Discriminative Knowledge Boosting Alignment (DKBA) for chest X-rays. The overall verdict is 'questionable' (yellow). The principal finding concerns selective reporting: Table 4 ablation results show that adding PAR to the baseline reduces ChestX-ray14 AUC from 79.1 to 78.9 (without DKBA) and from 81.0 to 80.8 (with DKBA), yet Section 5.3.1 claims PAR significantly improves performance and never acknowledges this regression. A secondary concern involves computational plausibility: 200 reconstruction pre-training epochs on MIMIC-CXR (~377,110 images and 227,835 reports) with ViT-B/16 plus BERT/PubMedBERT on only 2 RTX 4090 GPUs appears unrealistic, although no direct reproduction evidence is available. Additionally, Table 1 exhibits layout corruption, with Med-UniC's CXP/RSNA/COVIDx values concatenated into an unreadable string. Confidence in the selective-reporting finding is high; confidence in the compute claim is low and remains a reasonable suspicion. No definitive misconduct judgment is made.

Verdict

🟡 Questionable — The paper exhibits a clear inconsistency between its ablation table and its narrative claims, plus notable presentation issues, but no conclusive evidence of fabrication has been established.

Key findings

Selective / inconsistent reporting on PAR module (Orange severity, confirmed): Table 4 shows adding the Pathology-Aware Reconstruction (PAR) module decreases ChestX-ray14 AUC in both configurations (79.1 → 78.9 without DKBA; 81.0 → 80.8 with DKBA), yet Section 5.3.1 states PAR significantly improves fine-grained downstream performance without mentioning the regression on ChestX-ray14.
Compute plausibility concern (Yellow severity, insufficient evidence): 200 epochs of MIMIC-CXR pre-training with ViT-B/16 + BERT/PubMedBERT, including image reconstruction and graph attention operations, reportedly executed on only 2 NVIDIA RTX 4090 GPUs (24 GB each). No logs or timing data are provided, and the configuration is difficult to reconcile with standard experience; reproduction or training-log disclosure would be needed to verify.
Table layout corruption (Yellow severity, confirmed): In Table 1, the Med-UniC [34] ViT-based row collapses multiple cells into the continuous string "89.4 89.7 90.891.993.193.780.389.594.5", strongly suggesting a LaTeX/PDF extraction or compilation error in a camera-ready ACM MM '25 submission.

Evidence highlights

Table 4, ChestX-ray14 (AUC) column: baseline + DKBA = 81.0; +PAR +DKBA = 80.8 (Δ = −0.2). Without DKBA: baseline = 79.1; +PAR = 78.9 (Δ = −0.2).
Section 5.3.1 narrative claims PAR "significantly improves performance in fine-grained downstream tasks" without acknowledging the ChestX-ray14 AUC decrease.
Section 4.3 (Implementation Details) states pre-training on MIMIC-CXR (377,110 images / 227,835 reports) for 200 reconstruction epochs and 15 alignment epochs on 2× RTX 4090 GPUs.
Table 1 cell concatenation for Med-UniC [34] (ViT-based): "89.4 89.7 90.891.993.193.780.389.594.5" across CXP, RSNA, COVIDx columns.
DOI: 10.1145/3746027.3755336

Notes

The PAR negative result on ChestX-ray14 is reported in the original table, so this is primarily an issue of selective narrative emphasis rather than outright data fabrication. The most plausible benign explanation is that the authors emphasized gains on other datasets and omitted discussion of one unfavorable column.
The compute-plausibility concern is based on field experience with ViT-B/16 + BERT-scale multimodal pre-training on MIMIC-CXR; without access to training logs, hardware utilization traces, or wall-clock timings, it cannot be elevated beyond a reasonable suspicion.
The Table 1 layout issue is most likely a typesetting/PDF-extraction artifact, but it does undermine the rigor of the camera-ready artifact.
Recommended follow-up: (1) request authors release loss curves, wall-clock training times, and GPU utilization logs; (2) request a clarified discussion of PAR's ChestX-ray14 regression in an erratum; (3) request a corrected Table 1.
This is an AI-assisted review for academic discussion only; any formal misconduct determination requires institutional investigation.