How many specimens make a sufficient training set for automated 3D feature extraction?

Mulqueeney, J. M.; Searle-Barnes, A.; Brombacher, A.; Sweeney, M.; Goswami, A.; Ezard, T.

doi:10.1101/2024.01.10.575054

Home/Archive/bioRxiv

bioRxiv · 10.1101/2024.01.10.575054

How many specimens make a sufficient training set for automated 3D feature extraction?

Mulqueeney, J. M.·Searle-Barnes, A.·Brombacher, A.·Sweeney, M.·Goswami, A.·Ezard, T.

View original source ↗Export metadata ↓

Abstract

Deep learning has emerged as a robust tool for automating feature extraction from 3D images, offering an efficient alternative to labour-intensive and potentially biased manual image segmentation methods. However, there has been limited exploration into the optimal training set sizes, including assessing whether artificial expansion by data augmentation can achieve consistent results in less time and how consistent these benefits are across different types of traits. In this study, we manually segmented 50 planktonic foraminifera specimens from the genus Menardella to determine the minimum number of training images required to produce accurate volumetric and shape data from internal and external structures. The results reveal unsurprisingly that deep learning models improve with a larger number of training images and that data augmentation can enhance network accuracy by up to 8.0%. Notably, predicting both volumetric and shape measurements for the internal structure poses a greater challenge compared to the external structure, due to low contrast between different materials and increased geometric complexity. These results provide novel insight into optimal training set sizes for precise image segmentation of diverse traits and highlight the potential of data augmentation for enhancing multivariate feature extraction from 3D images. Subject CategoryLife Sciences - Computer Science Subject Areascomputational biology, Artificial Intelligence

Explore related subjects

evolutionary biology Preprint

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Mulqueeney, J. M., Searle-Barnes, A., Brombacher, A., Sweeney, M., Goswami, A., Ezard, T.. 2024-01-12. How many specimens make a sufficient training set for automated 3D feature extraction?. https://doi.org/10.1101/2024.01.10.575054

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

Ocean current models predict spatial structure for coastal biodiversity

Genomic diversity provides the raw material through which all evolutionary processes occur. Accurately describing how environmental features affect patterns of diversity shows promise in predicting the resilience of populations and entire communities to dynamic environments and novel stressors. Identifying the contribution of purely neutral responses to environmental effects can be difficult, especially across large spatial ranges with many covarying environmental features. Using an individual-based model of larvae carried by ocean currents, we predict patterns of neutral genomic diversity in a passively dispersed coastal species, suggesting large geographic regions of relative homogeneity with ancestry determined by the nearest "upstream" population with limited larval input. The genomic identity that dominates each region sharply transitions at the next such interruption, generating hierarchical structure and violating assumptions of isolation-by-distance. We call these upstream populations "lineage source regions" (LSRs), and we find strong concordance between LSR model predictions and an empirical population genomic survey of a keystone salt marsh mussel (Geukensia demissa). We quantitatively validate model performance and goodness-of-fit, updating baseline predictions of genomic diversity patterns from the classic isolation-by-distance model to this LSR framework. This spatially explicit coastal ancestry model has potential for generalization across passive dispersing species.

evolutionary biology↗

bioRxiv2026

Shared landscape, divergent genomes: Comparative population genomics of co-occurring flora in an isolated Australian mountain range

Understanding how landscape heterogeneity shapes genomic variation is key to conserving biodiversity and evolutionary potential. Yet conservation genomics is typically conducted at the single species level, limiting our ability to identify shared genomic patterns or understand how co-occurring taxa differ in their response to common environmental gradients. We assessed spatial patterns of genomic variation and their associations with geographic and environmental features among populations of nine co-existing plant species with contrasting life-histories and ecological traits across Gariwerd, a threatened biodiversity hotspot in south-eastern Australia. All species exhibited weak overall genetic structure, although the magnitude and spatial distribution of differentiation varied among species. Despite this variation, broadly consistent spatial trends were apparent, with higher genomic divergence largely concentrated in central zones and higher genomic diversity localised in southern regions of the landscape. Significant associations between genetic differentiation and geographic distance were observed in three species, while another species showed associations with environmental distance. These findings demonstrate that coexisting species can show overlapping spatial patterns in genomic variation while differing markedly in their associations with geographic and environmental variation across the same landscape. This comparative approach provides a basis for spatial prioritisation and conservation planning, while highlighting the complex eco-evolutionary processes operating across heterogeneous landscapes.

evolutionary biology↗

bioRxiv2026

BryoCAD: Extending Plant DNA Foundation Models Beyond Flowering Plants

Bryophytes exhibit diverse genome architectures and gene repertoires, yet their functional interpretation remains constrained by incomplete genomic annotations and limited experimental resources. Although DNA foundation models offer promising solutions, their applicability to evolutionarily divergent bryophytes remains largely unexplored. Here, we introduce BryoCAD, a DNA language model developed through continued pre-training of the angiosperm-specific PlantCAD2 model on 17 bryophyte genomes spanning hornworts, liverworts, and mosses. Across three representative bryophyte genomes and four gene-boundary motif recovery tasks, BryoCAD-Small improved mean zero-shot accuracy from 45.7% to 57.0%, a 24.8% relative increase over PlantCAD2-Small, while the over 200-fold larger Evo2-20B achieved a mean accuracy of only 39.3%. Remarkably, BryoCAD-Small outperformed PlantCAD2-Large in 11 of 12 motif-recovery comparisons across the three bryophytes, highlighting the benefits of lineage-specific adaptation over model scaling alone. These improvements extended to all six tested algal species, increasing mean accuracy from 26.3% to 37.5%, although performance varied across ferns and gymnosperms. BryoCAD also enhanced gene structure annotation: replacing PlantCAD2-Small with BryoCAD-Small increased transcript-level F1 scores from 0.578 to 0.625 in Marchantia polymorpha and from 0.485 to 0.530 in Physcomitrium patens. Together, our findings demonstrate that lineage-specific continued pre-training enhances genomic sequence interpretation and gene annotation across bryophytes, providing a transferable framework for extending DNA foundation models to evolutionarily diverse, understudied plant lineages.

evolutionary biology↗