What does this research mean for the field?

Training convolutional neural networks on reduced data fractions is not a reliable shortcut for architecture selection in biomedical imaging, as performance stability and architecture rankings are highly dataset-dependent and do not scale monotonically with data size. Novelty: ClaimNovelty.INCREMENTAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

To evaluate how training sample size affects overfitting, generalization stability, and architecture ranking consistency across convolutional neural networks in biomedical imaging tasks.

August 27, 2026BioengineeringOpen Access

CNN Sample-Size Effects Across Biomedical Datasets: A Reliability Pattern in Overfitting, Ranking, and Monotonicity

View Full Paper

Ask AI

Bookmark

Authors

GSGiacinto Angelo SgarroUniversity of Foggia MMMelle MendikowskiUniversität Hamburg DSDomenico SantoroMercatorum University

Discussion

Member takes

Implication

Benchmarking analysis uncovers four distinct reliability zones across biomedical image datasets, indicating reduced-sample screening fails to reliably identify optimal convolutional neural networks.

Key Points

To evaluate how training sample size affects overfitting, generalization stability, and architecture ranking consistency across convolutional neural networks in biomedical imaging tasks.
Trained 39 CNN architectures (1–3 layers, 16–64 filters) from scratch across 4 biomedical datasets: breast mammography, pediatric chest X-ray, brain tumor MRI, and skin lesion dermoscopy.
Evaluated performance across 7 training-set proportions (5%, 10%, 20%, 40%, 60%, 80%, and 100%) with 5 independent runs per configuration while maintaining a fixed test set size.
Revealed a dataset-dependent four-band reliability pattern: unstable below 20% data, uncertain overfitting between 20% and 60%, stable between 60% and 80%, and potentially counterproductive above 80%.
Demonstrated that CNN architecture rankings derived from reduced-data fractions do not reliably correlate with full-dataset rankings, undermining small-subset screening strategies.

Cite This Study

Sgarro et al. (2026) studied this question.

synapsesocial.com/papers/6a8fe98110c91c1e9262113f https://doi.org/10.3390/bioengineering13090971

View Full Paper

Ask AI

Bookmark

Implication

Benchmarking analysis uncovers four distinct reliability zones across biomedical image datasets, indicating reduced-sample screening fails to reliably identify optimal convolutional neural networks.

CNN Sample-Size Effects Across Biomedical Datasets: A Reliability Pattern in Overfitting, Ranking, and Monotonicity

Discussion

Member takes

Implication

Key Points

Cite This Study

CNN Sample-Size Effects Across Biomedical Datasets: A Reliability Pattern in Overfitting, Ranking, and Monotonicity

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider