PaperlayerView exact revision on arXiv

Source-linked AI summary

Protein Crystallization Segmentation and Classification Using Subordinate Color Channel in Fluorescence Microscopy Images

Truong X. Tran, Marc L. Pusey, Ramazan S. Aygun

arXiv:2610.15286v1cs.CVcs.AIcs.LG

TL;DR

Fluorescence dye can highlight crystals, but non-crystal reflections and emissions complicate analysis at the dominant wavelength. The paper tests percentile-based segmentation using a subordinate color channel at longer wavelengths. It reports lower misclassification and higher accuracy across four classifiers, including below-1% crystal misclassification in wet-lab experiments.

  • Problem

    Reflections and emissions from non-crystal regions can complicate fluorescence-microscopy analysis focused on a dye’s dominant emission wavelength.

    Source: §1 · p. 1

  • Method

    The paper compares subordinate-channel percentile-threshold segmentation with dominant-channel processing and evaluates extracted features using multiple classifiers.

    Source: §1 · p. 2, §7 · p. 15

  • Results

    Using the subordinate channel produced higher accuracy and lower MLC rates across four classifiers, with crystal misclassification below 1% reported in the conclusion.

    Source: §5.2 · p. 10, §7 · p. 15

  • Takeaways & Limitations

    The results indicate that subordinate-channel emission may be more helpful than the dominant channel for distinguishing crystals in TFL trial images.

    Source: §6 · p. 13, §7 · p. 15

    Some likely-lead and crystal images were still misclassified as non-crystals, particularly when images had low intensity or large bright backgrounds.

    Source: §6.2 · p. 14

Abstract

from arXiv · show

The accuracy of detecting protein crystals for fluorescence microscopy images is very critical for high throughput and automated systems. Although the trace fluorescent labeling method could highlight protein crystals, reflection and emission from the fluorescence dye is not always due to crystal regions. Therefore, the analysis of the peak wavelength in the emission spectra of a fluorophore may not always yield effective results. In this paper, we show that using the subordinate color intensity corresponding to longer wavelengths than the peak wavelength of the emission spectra could improve the accuracy of protein crystal detection. Hence, we have built a segmentation method based on the percentile intensity of the subordinate color for trace fluorescently labeled (TFL'd) protein crystallization trial images. Compared to using the dominant color channel, our segmentation method on subordinate color channel was able to reduce the misclassification rate of likely-leads or crystals as non-crystals by the percentage of from 9.71% to 2.02% depending on the classifier. Similarly, the accuracy of classifiers were increased by the percentage of from 1.77% to 5.53%. Our method reached around 94% accuracy while keeping misclassification of likely-leads and crystals as non-crystals below 1%. Moreover, to evaluate the generalizability of our method, we have conducted new wet lab experiments on two proteins, Concanavalin A (Con A) and Ab inorganic pyrophosphate (AbIPPase), and the misclassification rate was below 1%. Our experiments show that using the subordinate channel may be more helpful for TFL'd protein trial image classification.

1 Introduction

Fluorescence imaging can highlight protein crystals, but reflections and emissions from non-crystal regions complicate dominant-wavelength analysis. This paper tests subordinate-color segmentation against dominant-color methods and reports improved classification and wet-lab results.

1 Introduction

  • High-throughput crystallization systems acquire and classify many trial images, and TFL can make crystal-bearing wells visible under fluorescence illumination.Source: §1 · p. 1
  • Subordinate-color processing targets longer emission wavelengths, where non-crystal reflections may be lower while dye-labeled crystals can still emit significantly.The paper compares percentile-threshold features from subordinate and dominant channels across classifiers.Source: §1 · pp. 1–2
  • Across the classifiers, subordinate-channel features improved accuracy and reduced crystal or likely-lead misclassification as non-crystal compared with dominant-channel features.The introduction identifies this comparison as the paper’s focus.Source: §1 · p. 2
  • New wet-lab experiments on Concanavalin A and AbIPPase reported below-1% misclassification of crystals as non-crystals using the subordinate channel.Source: §1 · p. 2

2 Related Work

Prior crystallization-image studies used varied image features and classifiers, with reported performance spanning several classes and error measures. The authors identify a gap in fluorescence microscopy: comparison of dominant- and subordinate-color-channel classification.

2 Related Work

  • Earlier methods combined texture, edge, geometric, and wavelet features with neural networks, SVMs, decision trees, and reported varying accuracy and error rates.Spraggon et al. reported 47%–82% accuracy across categories; Zhu et al. reported a 9.6% lowest false-negative rate and 14.6% best false-positive rate.Source: §2 · pp. 2–3
  • Later work included multi-scale image features and random forests, while Bruno et al. reported 94% test-image recognition across four categories.Cumbaa et al. reported category accuracies of 80%, 89%, and 98%.Source: §2 · p. 3
  • TFL studies classified non-crystal, likely-lead, and crystal images, with reported results including 88% overall accuracy and, in later work, 96% accuracy with 87% sensitivity.The latter study used a large feature set including intensity, thresholding, histogram, and graph features.Source: §2 · p. 3
  • The authors note that fluorescence-microscopy studies had not compared classification using dominant versus subordinate color wavelengths.They also caution that methods are difficult to compare because studies vary in light source, features, and crystallization categories.Source: §2 · p. 3

3 Trace Fluorescent Labeling and Image Classification

The paper describes a TFL crystallization workflow, its 2,756-image dataset, dominant-channel thresholding, extracted image features, and the model-building process used to classify trial images.

3.1 Trace Fluorescent Labeling Protein Crystallization Analysis

  • TFL experiments label protein with fluorescent dye, prepare and incubate sample plates, periodically scan wells, then segment and classify images for crystallization scoring.Computer analysis results can be reviewed by the experimenter and used in later optimization experiments.Source: §3.1 · p. 4

3.2 TFL2756 data set

  • The TFL2756 dataset contains 2,756 expert-labeled images in non-crystal, likely-lead, and crystal categories; around 60% are non-crystals.Images are 320 × 240 pixels, 32-bit depth, and 96 dpi; likely-leads denote conditions that may lead to crystals despite lacking clear crystals.Source: §3.2 · p. 4

3.3 TFL’d Image Segmentation based on Dominant Color Wavelength

  • Because green is dominant under 530 nm excitation, the prior segmentation method thresholds green-channel percentiles and requires grayscale intensity above tmin.DomG90 denotes the 90th-percentile green threshold when green is the dominant channel; G100 selects pixels at maximum green intensity.Source: §3.3 · p. 5

3.4 Region Features Extraction

  • The feature set combines binary-image intensity and shape measures with properties of the five largest connected regions in each image.Seven binary-image features and nine features per selected region yield 52 features per image.Source: §3.4 · p. 5

3.5 Building Classification Model Using Image Segmentation Features

  • The classification framework trains and validates models on segmented-image features, tests them on held-out data, then trains a final model on all TFL2756 images.The final model is used to classify new wet-lab protein crystallization images.Source: §3.5 · p. 6

4 Proposed Method: Protein Crystallization Segmentation and Classification Using Subordinate Color Channel

The proposed method analyzes fluorescence trial images using a subordinate rather than dominant color channel. It uses longer-wavelength emission and percentile thresholding to segment the images.

4.1 Subordinate Color Channel in Fluorescence Emission

  • CR-6G emits at wavelengths above 590 nm despite peaking around 550 nm, with the 595–700 nm tail appearing in the red channel at lower relative intensity.The paper calls the longer-wavelength, lower-intensity color the subordinate channel.Source: §4.1 · pp. 6–7, Figure 8 · p. 8

4.2 Color Intensity Analysis of The Protein Crystallization Trial Images

  • In the example, subordinate red intensity separates non-crystal regions (around 34) from crystal regions (>225), while green intensity exceeds 210 in both.The passage presents these pixel values as an example of why red may better distinguish crystal from non-crystal regions.Source: §4.2 · p. 7, Figure 9 · p. 8

4.3 Subordinate Color Percentile Intensity Thresholding

  • SPI binarizes the subordinate red channel using a per-image percentile threshold, a color-specific tmin floor, and a tmax exception for pixels below the percentile threshold.The foreground rule requires intensity above both tmin and the percentile threshold, or intensity at least tmax.Source: §4.3 · pp. 7–8

5 Experiments and Results

The experiments compare subordinate-channel SPI segmentation with dominant-channel and overall-intensity thresholding across classifiers, threshold settings, validation schemes, and new wet-lab images. Subordinate-channel classification performed better across four classifiers, with around 94% average accuracy on TFL2756 and strong results on two new proteins.

5.1 Evaluation Measures

  • Evaluation prioritized misclassifying likely-leads or crystals as non-crystals, alongside crystal misses and overall accuracy; data were split 90/10 with category ratios preserved.The 90% training and 10% final-evaluation split was randomly stratified.Source: §5.1 · p. 10, §5 · p. 10

5.2 Performance of the Subordinate Color Channel on Classification

  • The comparison used Otsu overall-intensity thresholding, dominant-green percentiles, and subordinate-red percentiles, with 52 extracted attributes supplied to four classifiers.The subordinate-red settings included SubR90, SubR97, SubR99, and SubR100; the dominant-green settings included DomG90, DomG97, DomG99, and DomG100.Source: §5.2 · p. 10
  • Subordinate segmentation reduced MLC by 9.71% to 2.02% depending on classifier and raised classifier accuracy by 1.77% to 5.53%.The results favored subordinate-channel segmentation across NB, DT, RF, and MLP classifiers.Source: §5.2 · p. 10, Figure 14 · p. 11, Figure 16 · p. 11

5.3 Tunability of Thresholding Factors for the Subordinate Channel

5.4 Overall Performance using the Subordinate Color Channel

  • Around 94% average accuracy, with 0% MLC and 0% MC on 90/10 hold-out; cross-validation MLC was 0.57%, 0.43%, and 0.43%.All cross-validation tests had an MC rate of 0.62%.Source: §5.4 · p. 13, Table 1 · p. 13

5.5 Evaluation of the Subordinate Channel on New Wet-lab Experiments

  • On 384 new wet-lab trial images, accuracy was 90.36%, MLC was 1.42%, and MC was 0.60%.The images came from experiments on Concanavalin A and AbIPPase proteins.Source: §5.5 · p. 13

6 Discussion

The discussion relates subordinate-channel performance to fluorescence signals that can also arise outside crystals, and compares binarized images across thresholding methods. It also identifies residual misclassifications and discusses intensity-specific models and threshold choices.

6 Discussion

  • Fluorescence dye can emit at wavelengths beyond its peak, while free-floating dye or reflective chemicals can also produce non-crystal signals.The authors suggest that analyzing wavelengths beyond the illumination color may increase robustness.Source: §6 · pp. 13–14

6.1 Comparing Binarized Images of Dominant and Subordinate Color Channels

  • SubR99 produced the best visual segmentation among the compared methods on difficult images, removing background regions and retaining clear crystal regions.Otsu included background; DomG90 improved slightly, while DomG99, SubR90, and SubR99 performed better in the illustrated cases.Source: §6.1 · p. 14, Figure 22 · p. 14, Figure 23 · p. 14

6.2 Misclassified Trial Images

  • Some likely-leads and crystals were still misclassified as non-crystals, particularly with low-intensity images or large bright backgrounds.The authors suggest proper lighting and in-focus images as ways to improve image quality.Source: §6.2 · p. 14

6.3 Building Multiple Models for Varying Intensities.

  • For images with varying intensities, the authors propose several models for intensity ranges and selecting the corresponding model for each input image.They note this may help when lighting problems make captured images differ from usual intensities.Source: §6.3 · pp. 14–15

6.4 Evaluation of Thresholds

  • SubR99 MLC fell as tmin decreased, while different tmax values gave the same MLC at a fixed tmin; the authors suggest tmin near 0 and retaining tmax.They suggest the method may use tmin and the percentile threshold, TCP, with the same performance.Source: §6.4 · p. 15

7 Conclusion

The paper introduces subordinate-channel segmentation for classifying trace-fluorescently-labeled protein crystallization images. Results favor this channel over the dominant wavelength across four classifiers, with misclassification of crystals below 1%.

7 Conclusion

  • Subordinate-channel segmentation reduced crystal misclassification below 1% and produced higher accuracy and lower MLC rates across four classifiers.The authors also report that subordinate-channel segmentation benefited from varying threshold values.Source: §7 · p. 15

Author Biography

The biographies outline the authors’ academic backgrounds, current roles, and research interests.

Author Biography

  • Tran is pursuing a doctorate in computer science at the University of Alabama in Huntsville and works on protein crystallization research data analysis.His research interests include machine learning and data mining.Source: Author Biography · p. 16
  • Pusey is a research scientist at iXpressGenes focused on improving protein crystal screening, detection, and crystallization-condition identification using visible fluorescence.He previously worked on protein crystal growth at NASA/MSFC.Source: Author Biography · p. 16
  • Aygun is an associate professor in computer science at the University of Alabama in Huntsville whose interests include protein crystallization image analysis and data mining.His interests also span image and video processing, multimedia databases, and related areas.Source: Author Biography · pp. 16–17

Related papers

Loading 2610.15286v1…