Academic Integrity Investigation Report: 'Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training' (DOI: 10.1145/3746027.3755336)

Academic fraud report · Geng Detector

Summary

This report assesses a paper published at ACM Multimedia 2025 (DOI: 10.1145/3746027.3755336) and returns a verdict of 'highly suspicious.' The central concerns involve implausible computational claims: the authors state they pre-trained on the MIMIC-CXR dataset (≈377k images and 228k reports) for 200 epochs of visual reconstruction plus 15 epochs of alignment using only two NVIDIA RTX 4090 GPUs (24GB each), despite peer works typically requiring 8+ A100 (80GB) GPUs for weeks or months. A second red flag is Table 1, where the proposed model achieves an identical AUC of 89.5 on CheXpert (CXP) at both 1% and 10% training data, rising only marginally to 89.7 at 100%—a near-impossible outcome for real experiments. Third, the ablation in Table 4 shows the headline PAR module either regressing (AUC 79.1→78.9 on ChestX-ray14) or contributing negligibly on core classification tasks, contradicting the paper's narrative. Additional formatting artifacts (number concatenation, text extraction errors) suggest hasty data fabrication. Image-based checks were not feasible. Overall confidence in the identified textual anomalies is high; verification of visual figures and training logs is required to escalate the finding.

Verdict

Highly suspicious (🟠). Multiple independent anomalies cluster around resource claims, reported metrics, and ablation behavior. Confirmation requires access to training logs, code, and original figures.

Key findings

Implausible compute claim: 200 epochs + 15 epochs of multimodal pre-training on MIMIC-CXR (≈377k images, 228k reports) reportedly performed on only 2× RTX 4090 (24GB). Comparable peer works at this scale typically require 8+ A100 (80GB) GPUs over weeks to months (Section 4.3 Implementation Details).
Implausibly identical results: On CheXpert (CXP) with ViT-based evaluation (Table 1), the proposed model reports AUC = 89.5 at 1% data, AUC = 89.5 at 10% data, and AUC = 89.7 at 100% data. A 10× increase in training data producing no measurable change is inconsistent with real deep-learning experiments.
Inconsistent ablation: The PAR module, half of the paper's title and stated core contribution, causes AUC to drop (79.1→78.9) on ChestX-ray14 and produces negligible ACC gain on COVIDx (77.0→77.5) when added in isolation (Table 4 / Section 5.3.1). Improvements appear only in the combined configuration, suggesting post-hoc narrative construction.
Document/formatting artifacts: Number strings concatenated without separators in Table 1 (e.g., 90.891.993.193.780.389.594.5) and extraction garble such as &#3627408451; indicate data may not originate from authentic experiment logs.
Image-level analysis could not be performed in this report; Figure 5 (t-SNE) and Figure 6 (attention maps) warrant original-file inspection.

Evidence highlights

DOI: 10.1145/3746027.3755336 (ACM MM '25, October 27–31, 2025).
Hardware statement: "two NVIDIA RTX 4090 GPU" combined with 200 reconstruction epochs + 15 alignment epochs on MIMIC-CXR.
Table 1 (CXP, ViT-based, Ours): 1% → 89.5; 10% → 89.5; 100% → 89.7.
Table 4 ablation (PAR only): ChestX-ray14 AUC 79.1 → 78.9; COVIDx ACC 77.0 → 77.5.
Stated code repository: https://github.com/Felix1118/PADKB (not independently verified here).

Notes

Confidence is high for the textual and tabular anomalies; low for image manipulation due to lack of original pixel data.
PDF text-extraction artifacts could partially explain formatting issues, but the numeric duplication in Table 1 is a substantive, not cosmetic, concern.
Recommended follow-up: request raw training logs and W&B traces, audit batch size and memory usage in the released code, post structured queries on PubPeer, and notify the ACM MM 2025 program committee.
This is an AI-assisted assessment for academic discussion only; any formal finding of misconduct requires institutional investigation.