cs.CVSep 19, 2026

Reconstructed holograms and explanation-aware evaluation for low-cost computational pollen analysis in veterinary cytology

Authors: Swarn Warshaneyan, Joial Danyal, Blaž Cugmas, Mindaugas Tamošiūnas, Edgars Kviesis-Kipge, Kirishanth Manivannan, Roberts Kadiķis

Organizations: Institute of Electronics and Computer Science (EDI), Riga, Latvia · Faculty of Science and Technology, University of Latvia, Riga, Latvia

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Automated pollen analysis supports veterinary cytology, but brightfield microscopy is costlier and more complex than lens-less digital in-line holographic microscopy. We evaluate whether reconstructed holograms can narrow this gap and whether model explanations remain reliable under modality change. Six pollen species were imaged by brightfield and holographic microscopy. Raw, single back-propagation and iterative phase retrieval holograms were evaluated with YOLOv26s detection and MobileNetV4 classification after anchor-based annotation transfer. Six attribution methods were assessed for spatial grounding and faithfulness with the Attribution Health Inspection and Repair (AHIR) protocol, which tests model brittleness under weak noise and corrects attribution-map granularity when needed. Brightfield achieved 0.6890 mAP50-95 (0.8865 mAP50) for detection and 0.9687 macro-F1 (0.9705 accuracy) for classification. Reconstructed holograms narrowed the gap with a task-dependent split: p-type was strongest for detection at 0.5324 mAP50-95 (0.8229 mAP50), while r-type was strongest for classification at 0.7695 macro-F1 (0.7866 accuracy), both far above raw-hologram baselines. Activation-based explanations localized strongly on grains, and region-based methods retained ~60 to ~80% of faithfulness under holography. The holographic detector was highly brittle to weak perturbations, saturating deletion-based evaluation while insertion remained informative. Pixel-level gradient explanations approached random floor, yet spatial smoothing restored p-type gradient faithfulness from 0.05 to 0.51. For holographic classification, perturbation-based explanations remained faithful while gradient-based methods fell below random floor. Reconstruction improves low-cost holographic pollen analysis, while AHIR distinguishes genuine attribution failure from artifacts caused by model brittleness and map granularity.

Figures & tables

Figure 1: Imaging setups for the two strict acquisition modalities. (a) Optical acquisition arm of the inverted brightfield microscope (Nikon Ti2-E). (b) Lens-less DIHM setup schematic with an example hologram, reproduced from [ 9 ] under the CC BY 4.0 license. (c) Lens-less acquisition chain used for this dataset, after [ 9 ] .
Figure 2: Fixation methods compared on raw holograms. A tape-fixated raw hologram from the previous dataset [ 50 ] and a silicone-fixated raw hologram from the current dataset. The silicone preparation yields a visually cleaner field.
Figure 3: The four effective imaging modalities at grain scale. Zoomed views of the same field for optical/brightfield, raw hologram, p-type (single-backpropagation) and r-type (Gerchberg-Saxton iterative) reconstructions; whole-field views are given in Fig S5.
Figure 4: Anchor-based registration of the holographic modalities to the optical reference. Left: scale-invariant feature transform (SIFT) + RANSAC keypoint matches between the raw hologram and the brightfield reference under the reused affine transform. Right: consistency check between the raw hologram and the r-type reconstruction at identical coordinates. The black band on the right edge of the left panel is the out-of-frame region produced by the size mismatch between the modalities. The r-type reconstruction is registered to the optical image. The resulting partial-affine transform is reused for the raw and p-type holograms.
Figure 5: Narrowing of the optical-holographic gap relative to the previous dataset. Best previously published raw-hologram values (purple; detection includes synthetic-data augmentation [ 50 ] ) against the current results at 100% bounding box area. The two datasets differ in fixation, acquisition, and preparation, so the comparison reflects the progress of the overall pipeline rather than a controlled benchmark.
Figure 6: Representative Layer-CAM-G saliency maps, optical versus p-type. Layer-CAM-G attribution overlaid on (a) an optical patch and (b) a p-type reconstruction patch for the same species. Maps are computed per detection within 640-pixel patches, and the attribution concentrates on the target grain in both modalities.
Figure 7: IxG attribution maps before and after spatial smoothing. (a) An image patch, (b) its pixel-scattered plain Input ×\times Gradient map, and (c) the same map after spatial smoothing. Smoothing turns a correct but fragmented ranking into region-scale structure.
Figure 8: Granularity rescue is method-specific. Over-all score of the plain, smoothed ( σ\sigma = 4, 8, 16) and random variants of pixel-level IxG maps and region-level plain Layer-CAM maps on the same matched detections, (a) optical and (b) p-type; error bars are 95% bootstrap confidence intervals.
Figure 9: Matched-grain saliency maps for the two reconstructions. Layer-CAM-G attribution overlaid on the p-type (top row) and r-type (bottom row) patches for the same physical grains detected at identical coordinates; two high-agreement pairs and one low-agreement pair are shown with their map rank correlations. Panels are full 640-pixel patches.
Figure 10: Noise-gate dose-response. (a) Median retention of the matched detection score with interquartile range and (b) fraction of fully collapsed detections as a function of the noise standard deviation; the dashed line marks the operating point of 8/255.
Figure 11: Deletion- versus insertion-side detector metrics. D-Deletion and D-Insertion ( γ\gamma = 0.45) for random, plain and smoothed maps on (a) optical and (b) p-type detections; error bars are 95% bootstrap intervals.
Figure S1: Bounding box areas of 100%, 125%, and 150% on a raw hologram.
Figure S2: Size gap between the modalities after alignment. The raw hologram aligned onto the brightfield reference; the surrounding margin shows how much smaller the holographic field of view is. Warped regions falling outside the holographic frame become black borders, and crops containing them were removed.
Figure S3: Per-species Occlusion detection explanation faithfulness. Over-all score per pollen species for optical and p-type images.
Figure S4: Per-species classification explanation faithfulness on r-type crops. Over-all score per pollen species and attribution method.
Figure S5: Whole-field views of the four effective imaging modalities for a single pollen sample. Optical/brightfield, raw hologram, p-type and r-type reconstructions of the same field of view; the zoomed views are Fig 3 of the main text.
Figure S6: Localization lift by method and modality. In-box energy as a multiple of geometric chance (log scale); the dashed line marks the random-map level.
Armn2/2554/2558/25516/25532/255
Detection, optical3720.992 (0%)0.954 (4%)0.706 (18%)0.000 (64%)0.000 (100%)
Detection, p-type3060.753 (16%)0.000 (100%)0.000 (100%)0.000 (100%)0.000 (100%)
Table S8: Noise-gate dose response. Median score retention and fraction of items with zero retention under Gaussian pixel noise, five draws per item, for the detection arms; the operating point of the main text is σ\sigma = 8/255.
MapD-Del. opticalD-Ins. opticalD-Del. p-typeD-Ins. p-type
Random0.1160.1400.0110.011
IxG0.0190.6040.0100.077
IxG, smoothed0.0130.7080.0090.518
Layer-CAM0.0190.6680.0410.477
Layer-CAM, smoothed0.0130.7460.0090.607
Table S9: D-Deletion and D-Insertion at γ\gamma = 0.50. Means over the same 376 optical and 303 p-type matched detections as the γ\gamma = 0.45 values of the main text; the ordering of maps is unchanged.
Armn passOA, passn failOA, fail
Detection, optical228+0.674 [+0.650, +0.697]144+0.442 [+0.408, +0.474]
Detection, p-type0no items306+0.054 [+0.049, +0.060]
Classification, optical87+0.050 [+0.035, +0.065]203-0.001 [-0.009, +0.007]
Classification, r-type148-0.024 [-0.033, -0.014]135-0.006 [-0.012, -0.001]
Table S11: Gate validation across evaluated arms. Over-all score of the plain IxG map for items that pass the noise gate (retention ≥\geq 0.5) and items that fail it, with 95% bootstrap intervals.
More in this topic

Microscopy Image Analysis

Microscopy image analysis uses computational methods to detect, segment, track, reconstruct, and quantify structures in scientific images.

See all 198 papers

Explore similar work

CardsList
Jun 16, 2026

Million-scale multimodal pollen microscopy with expert-guided foundation models

Automated pollen identification from microscopy remains a bottleneck in aerobiology, palaeoecology and biodiversity monitoring, because scalable systems must generalise across specimen preparation, scanner settings and geographic origins while retaining palynological interpretability. To address this gap, we present a million-scale multimodal pollen microscopy resource, Pollen AI Atlas, assembled from pure-species whole-slide bright-field images spanning four geographic origins, four scanner settings and 46 taxon labels across 31 botanical families. Seeded by one manually selected exemplar per source slide, token-level mining and filtering produced 1,511,390 released grain detections with 99.6% proposal precision in expert-curated test regions. Each detection was paired with machine-generated grain-level morphological captions from five open-weight vision-language models, guided by expert-verified palynological anchors, yielding structured descriptions of aperture systems, wall ornamentation, shape and size. Among the evaluated models, Gemma4 provided the most controlled primary caption set, combining tight length control, no leakage and the strongest text-retrieval performance. Baseline benchmarks with frozen visual features reached 88.16% top-1 accuracy, while cross-regional retrieval showed that caption-derived text embeddings remained robust when image similarity degraded (mAP@20 0.811 versus 0.262). Released data, annotations, captions, splits, code, and weights provide a benchmark for pollen recognition, cross-regional domain adaptation and domain-specific multimodal microscopy learning.

András Biricz, Björn Gedda, Donát Magyar +4

Department of Physics of Complex Systems, ELTE Eötvös Lor´and University, Budapest, Hungary · 2The Palynological Laboratory at the Swedish Museum of Natural History, Stockholm, Sweden · 3National Centre for Public Health and Pharmacy, Budapest, Hungary +4

Domain AdaptationMultimodal Model EvaluationMicroscopy Image Analysis
AskPDF
Apr 17, 2026

Automated Palynological Analysis System: Integrating Deep Metric Learning, Detection and Classification in Bright Field Microscopy

Traditional melissopalynology is a time-consuming and subjective process, often taking 4-6 hours per sample. We present an automated, high-throughput microscopy system that integrates H_\infty robust mechanical control with advanced deep learning pipelines for the precise counting, classification, and morphological analysis of pollen grains from Bio Bio region in south central territory in Chile. Our system employs U^2-Net for salient object detection and a DINOv2 Vision Transformer backbone trained via Deep Metric Learning for classification. By integrating Gradient-Weighted Attention, the model provides human-interpretable texture and diagnostic feature annotations. The system achieves a 95.8% classification recall and at least 6x processing speedup compared to manual expert analysis.

J. Staforelli-Vivanco, R. Jofré, P. Coelho +12

Facultad de Ingeniería, Universidad Andres Bello, Talcahuano, CCP-THNO 7100 Chile · Universidad Arturo Prat, Sede Victoria, Chile · Laboratory of Palynology and Plant Ecology, School of Sciences and Technologies, University of Concepción, Los Angeles Campus

Salient Object DetectionExplainable Image ClassificationMicroscopy Image AnalysisImage Classification
AskPDF
Sep 10, 2026

Hologram Representation via Quadratic Phase Gaussian Splatting

We introduce Complex-Valued Quadratic Phase Gaussian (CVQPG), a novel hologram representation method that augments each 2D Gaussian primitive with a quadratic phase profile controlled by a learnable curvature parameter. Against the planar Gaussian baseline, CVQPG improves the average PSNR of holographic reconstructions by 0.19 dB (RGB) and 0.33 dB (grayscale) at equal primitive counts, and by 0.05 dB (RGB) and 0.08 dB (grayscale) at equal parameter counts, where it still leads in all visual quality metrics. Our frequency-domain analysis shows that CVQPG better preserves the mid-to-high frequency band of natural images, where the reconstruction MSE drops by up to 11% (RGB) and 22% (grayscale), indicating that modulating primitive wavefronts is an effective and lightweight enhancement.

Haolong Wang, Yicheng Zhan, Kaan Akşit +1

Swansea University, United Kingdom · University College London, United Kingdom

Image Reconstruction3D Scene RepresentationGaussian Splatting
AskPDF