Methodological audit and empirical evaluation of a lightweight whole-volume 3D CNN for preterm brain injury classification

This study uses methodological auditing to evaluate a proof-of-concept engineering pipeline for classifying radiological severity groups of preterm brain injury (BI) from T2-weighted MRI in the dHCP cohort. Its primary contribution is methodology-driven empirical discovery rather than a new neural-network architecture. GloResNet combines a standard ResNet-10 backbone, MedicalNet initialization, whole-volume array resizing, mixup, and array-axis-0-reversal probability averaging, hereafter referred to as axis-0-reversal averaging; no new CNN block is proposed.

The audit identifies concrete mechanisms by which small-cohort imaging results can be overstated, including unit-of-analysis drift, fold-assignment non-traceability, post hoc model-selection bias, and ranking across incompatible evaluation protocols. Building on this framework, we report the first fair, traceable comparison between a de novo trained Vision Transformer (ViT) and a domain-adapted lightweight 3D CNN on a public preterm brain injury dataset. When all sources of bias are controlled, the lightweight CNN achieves comparable or superior performance to ViT, challenging the prevailing assumption that larger, more complex architectures are always preferable in low-data regimes.

Applying the audit to the historical evidence also showed that the archived GloResNet mean fold accuracy of \(69.12 \pm 7.42\%\) was calculated over 180 scan records from 128 subjects, including repeated scans, without retained subject identifiers or out-of-fold predictions. The historical archive is therefore reported separately from the fair, traceable server-side comparison. A separate, auditable subject-level input-representation experiment did not demonstrate superiority of whole-volume input.

The resulting minimum reporting standard requires an explicit statistical unit, traceable subject-level folds, retained out-of-fold predictions, prespecified selection rules, and transparent comparator conditions. These internal estimates do not establish clinical screening performance or external generalizability. The source code is available at https://github.com/ICL-SUST/GloResNet-Preterm-Brain We thank the Developing Human Connectome Project for providing the neonatal MRI dataset.

Declaration of Generative AI and AI-assisted technologies in the manuscript preparation process: During the preparation of this work the authors used Gemini and ChatGPT in order to polish the English language and organize foundational data. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article. This work was supported by the Shenzhen Science and Technology Program (Grant Nos.

JCYJ20240813152044056 and JCYJ20250604185306008) and the Sanming Project of Medicine in Shenzhen Guangming (No. szgmtd2025001). Image Computing Laboratory, Shaanxi University of Science and Technology, Xi’an, China Department of Neonatology, Shenzhen University of Advanced Technology General Hospital, Shenzhen, China Department of Neurosurgery, The First Affiliated Hospital of Xi’an Jiaotong University, Xi’an, China CSIRO Technology, PO Box 76, 1710, Epping, NSW, Australia Correspondence to Weichuan Zhang, Benqing Wu, Tuo Wang, Changming Sun or Liang Guo. This study was a secondary analysis of de-identified, publicly released dHCP data.