Source-linked AI summary
Scalable Detection of Fossil Palynomorphs in Multifocal Digital Microscopy Images
Abbas Shaikh, Praise Mayor, Patrick Ainlay-Vazquez, Aditya Viswanathan, Teon Golden, Eric Zhang, Ingrid C. Romero, Alexander E. White, Scott Wing, Arko Barman
TL;DR
The paper addresses the difficulty of scaling palynomorph analysis from manual inspection of massive multifocal slides to automated whole-slide detection. It builds a tiled, compressed, end-to-end pipeline with modern detectors, efficient detection synthesis, and optimized I/O, achieving strong detection performance and under-one-hour inference.
Problem
Manual palynomorph analysis is slow, while multifocal whole-slide images are too large and complex for straightforward standard object-detection processing.
Method
The pipeline tiles slides, compresses multifocal stacks into 2D inputs, benchmarks YOLO26-L and RF-DETR-2XL, merges overlapping-tile detections efficiently, and optimizes I/O.
Results
AP@50 0.879 and AP@50-95 0.642 were achieved by the best RF-DETR configuration, while whole-slide detection time was reduced to under one hour.
Takeaways & Limitations
The pipeline supports large-scale automated analysis of multifocal palynomorph slides and is designed to extend to other large-scale microscopy applications.
Takeaways & Limitations
The evaluation used 82 annotated slides, and the authors report uneven difficulty across splits that may limit the reliability of generalization estimates.
Abstract
from arXiv · showhide
Palynomorphs (microscopic, organic-walled fossils such as pollen, spores, and dinoflagellates) are important high-resolution records of past climates and are critical to the study of ancient ecosystems. Existing methods rely on manual analysis of high-resolution, multifocal digital microscopy images, which is slow and time-consuming and requires researchers to compromise on the scale of their investigations. To the best of our knowledge, our work proposes the first ever scalable end-to-end pipeline for automated palynomorph detection in whole slide images that addresses this bottleneck through: (1) efficient methods for decomposing and compressing digitized multifocal microscope slide images into tractable 2-dimensional tiles for analysis; (2) benchmarking modern object detection models, including RF-DETR, for the detection of palynomorphs, achieving an AP@50 of 0.879; (3) an efficient algorithm for the synthesis of detection outputs across large-scale, high-resolution images; and (4) an I/O optimization resulting in faster inference time. Our methods drastically reduce the time required for palynomorph detection in a single slide from often days of manual inspection to under one hour of automated analysis, enabling palynological research at a substantially greater scale.
1. Introduction
Palynomorph slides are scientifically valuable but difficult to analyze manually at scale because specimens are dense, multifocal, and embedded in massive whole-slide images. Existing detection work has not resolved whole-slide, multifocal throughput, motivating an end-to-end scalable pipeline.
- Palynomorphs are well-preserved fossil particles that provide records of past environments and inform climate reconstruction.
- Manual analysis is slow because palynomorphs are interspersed with other sediments, making them difficult to distinguish and localize quickly.
- Multifocal slides can exceed 25 GB, making whole-image processing computationally infeasible with standard object-detection preprocessing, modeling, and post-processing.
- Three-dimensional specimens, occlusion, clutter, and preparation-induced appearance shifts make detection difficult and limit assumed transfer across slides.
- Prior automated work largely used cropped specimens or small fields of view, while whole-slide multifocal detection remained unaddressed.
- The proposed pipeline combines multifocal preprocessing, modern detectors, scalable output synthesis, and I/O optimization for operational-scale inference.
2. Related Work
Related work has emphasized classification and CNN-based localization of cropped pollen or small microscopy fields. Transformer-based detection has appeared in adjacent pollen settings but remains relatively unexplored for scalable palynomorph detection.
- Earlier computational palynology primarily classified pre-cropped specimen images rather than detecting objects across whole slides.
- Prior object-detection studies used YOLO, Faster R-CNN, RetinaNet, and specialized CNN methods for pollen localization and recognition.
- Transformer-based architectures have been less explored in palynology, with prior examples using vision transformers or RT-DETR in pollen-related settings.
3. Data
The dataset contains large, multifocal microscopy slides spanning diverse collection and preparation conditions. Only a curated subset was annotated, while focal-plane variation motivates preprocessing before two-dimensional detection.
- The dataset contains 847 Smithsonian slides from North American sites, with 25–27 focal planes and slide sizes exceeding 25 GB.
- Among these slides, 82 were selected for annotation and curation using rectangular regions of interest.
- The images vary in staining, mounting media, palynomorph density, subtype, and organic or inorganic matter across diverse environments.
- Annotated regions show substantial diversity and scale of organic and inorganic matter, underscoring the need for automated analysis.
- Across a 25-plane z-stack, image sharpness varies by focal plane, and no single global plane typically makes all objects clear.
4. Methods
The methods convert massive multifocal slides into overlapping two-dimensional tiles, benchmark two detectors under two focal-plane compression strategies, and synthesize detections efficiently across tiles. I/O changes further reduce the cost of whole-slide inference without changing outputs.
- Pre-processing: The pipeline tiles high-resolution slides, compresses focal stacks into two-dimensional images, detects palynomorphs, and synthesizes complete whole-slide results.
- Pre-processing: Annotated regions are divided into 1024×1024 tiles with 10% overlap to preserve high resolution and capture boundary-spanning specimens.
- Pre-processing: Ground-truth annotations are mapped to pixels, converted to rectangular boxes, filtered at tile boundaries, and stored with focal-plane dimensions.
- Focal-plane compression: Focus stacking selects each pixel from the focal plane with the maximum Laplacian-of-Gaussian response, while focal-plane selection provides an alternative compression method.
- Detection models: The study frames detection as a single-class task and benchmarks YOLO26-L against transformer-based RF-DETR-2XL with augmentation for collection variability.
- Inference optimization: I/O optimization addresses tile-reading bottlenecks through caching and related input-path changes, reducing input cost from approximately 3 s to 0.2 s per tile while preserving outputs.
- Post-processing: Boundary-aware de-duplication compares detections only across overlapping tiles and uses union-find to retain the highest-scoring detection in each connected component.
5. Experiments and Results
Experiments compare model architectures and multifocal-image compression methods, then evaluate qualitative detections and whole-slide inference efficiency. RF-DETR-2XL with focus-stacked inputs provides the strongest reported detection performance, while optimized I/O reduces production inference to under an hour per slide.
- Experiment design: The 82 annotated slides were split into 58 training, 12 validation, and 12 held-out test slides, preventing slide-level data leakage.The split was 70/15/15, with comparable annotation density, subtype distribution, and staining method across splits.
- Model and preprocessing comparison: RF-DETR-2XL outperforms YOLO26-L for both focal plane compression methods, with AP@50 margins of 0.053 for focus stacking and 0.017 for focal plane selection.RF-DETR-2XL also achieves larger recall margins, while the architectures respond differently to compression choice.
- Model and preprocessing comparison: 0.879 AP@50 and 0.642 AP@50-95 are achieved by RF-DETR-2XL on focus-stacked images, the best overall configuration.This configuration is used as the final model for evaluating I/O optimization methods.
- Qualitative results: RF-DETR-2XL recovers more true positives in dense, overlapping regions, whereas YOLO26-L misses specimens more frequently in cluttered areas.The qualitative comparison uses a representative held-out test tile with focus-stacked inputs.
- Whole-slide inference: 51 minutes per slide is achieved after I/O optimization for RF-DETR-2XL whole-slide inference, roughly an order of magnitude faster end-to-end.Reading and focal compression take 0.18–0.24 seconds per tile, while inference takes 0.15–0.20 seconds per tile using four CPU workers and one GPU.
- Operating threshold: τ⋆ = 0.404 yields precision 0.824, recall 0.810, and F1 = 0.817 for the best overall configuration.Because the threshold is selected on the same data used for performance reporting, it represents a best-case operating point rather than an unbiased deployment estimate.
6. Discussion
RF-DETR provided the strongest detection performance, particularly for recall, while focus-stacked inputs worked well despite their artifacts. The pipeline remains limited by sampling, discarded volumetric information, and the absence of subtype identification.
- Model Performance: RF-DETR-2XL achieved the best inference performance across precision, recall, AP@50, and AP@50-95, outperforming the CNN baseline.The model was trained and evaluated on focus-stacked images.
- Model Performance: RF-DETR reduced false negatives relative to YOLO26, although it produced occasional false positives in highly occluded regions.The authors attribute the difference partly to RF-DETR’s one-to-one matching strategy.
- Model Performance: Higher recall is preferred because experts can remove false positives, whereas false negatives represent undetected specimens.Thresholds can be adjusted according to review capacity or desired data completeness.
- Comparison of Focal Plane Compression Methods: Focal plane selection improved YOLO26 over focus stacking but provided no discernible benefit for RF-DETR, whose results were nearly identical across compression strategies.Focus stacking can introduce halo edges, blending seams, and other high-frequency artifacts.
- Comparison of Focal Plane Compression Methods: RF-DETR appeared robust to localized focus-stacking artifacts and could leverage sharpness changes that may be lost in a single representative focal plane.The authors also suggest aggressive augmentation may have reduced focus stacking’s sharpness advantage.
- Limitations and Future Work: The 82-slide sample may provide unreliable generalization estimates because splits may lack sufficient stratification of palynomorph subtypes and slide characteristics.The authors observed a consistent validation–test performance gap.
- Limitations and Future Work: Focus stacking and focal plane selection discard volumetric information, while raw multifocal models may retain depth at substantially higher inference cost.The paper presents this trade-off as a direction for future investigation.
- Limitations and Future Work: The current detection system does not identify palynomorph subtypes, which require subtle taxonomic distinctions.Subtype identification is proposed as future work.
7. Conclusion
The study presents an end-to-end pipeline for automated palynomorph detection in multifocal whole-slide images. RF-DETR achieved strong detection metrics, while I/O optimization reduced detection time to under one hour and supported scalable analysis.
- Conclusion: The pipeline integrates structured data handling, I/O optimization, depth-aware preprocessing, modern detectors, and boundary-aware NMS for multifocal whole-slide images.It is designed to address the scale and complexity of high-resolution microscopy data.
- Conclusion: RF-DETR achieved AP@50 of 0.879 and AP@50-95 of 0.642 while capturing complex visual patterns in palynomorphs.These metrics summarize the reported detection performance.
- Conclusion: The pipeline reduced detection time to under one hour and incorporated multifocal information for dense, heterogeneous images.The paper presents this as evidence of efficiency and scalability.
- Conclusion: The framework is intended to support more comprehensive and efficient reconstruction of past environmental conditions and Earth’s climate history.The stated consequence is framed within palynology and paleontology.