Methodological audit and empirical evaluation of a lightweight whole-volume 3D CNN for preterm brain injury classification
Abstract
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 ± 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
Related
- Efficient and real-time polyp segmentation: An empirical evaluation of lightweight Mamba integration strategies
- Pain, cognition, and culture in dementia research: a critical synthesis of empirical and methodological literature
- DSR-VGG: A Lightweight Depthwise Separable Residual VGG Architecture for Nasopharyngeal Histopathological Image Classification
- Scaled m-BonacciNet: A Recurrence-Guided Lightweight CNN for Efficient Medical Image Classification
- Psychometric evaluation of the Persian Version of the Postnatal Care Experience Scale (PCES) among Iranian postpartum women: a cross-sectional methodological study
BibTeX
@article{yuan2026methodological,
title = {{Methodological audit and empirical evaluation of a lightweight whole-volume 3D CNN for preterm brain injury classification}},
author = {Boyu Yuan and Jiamiao Lu and Weichuan Zhang and Benqing Wu and Tuo Wang and Changshan Wang and Changming Sun and Liang Guo},
journal = {Scientific Reports},
year = {2026},
month = oct,
doi = {10.1038/s41598-026-77358-z},
url = {https://www.nature.com/articles/s41598-026-77358-z},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Methodological audit and empirical evaluation of a lightweight whole-volume 3D CNN for preterm brain injury classification
AU - Boyu Yuan
AU - Jiamiao Lu
AU - Weichuan Zhang
AU - Benqing Wu
AU - Tuo Wang
AU - Changshan Wang
AU - Changming Sun
AU - Liang Guo
PY - 2026
DA - 2026/10/23
JO - Scientific Reports
DO - 10.1038/s41598-026-77358-z
UR - https://www.nature.com/articles/s41598-026-77358-z
AB - 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 ± 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
SN - 2045-2322
ER -