A Comprehensive Review of Image Analysis Methods for Microorganism Counting: From Classical Image Processing to Deep Learning Approaches

1.1 Basic knowledge of microorganisms

• Microorganisms include bacteria, viruses, fungi, and algae, with beneficial applications and harmful effects across human health, industry, food, and environmental systems.Examples include penicillin, fermentation, waste treatment, disease, epidemics, and ecological damage from algal blooms.Source: §1.1 · pp. 2–3
• Traditional counting methods are condition-dependent, laborious, subjective, and error-prone, whereas image analysis can count particles automatically, quickly, and objectively.Plate counting requires culturing and can undercount aggregated cells; hemocytometry cannot distinguish live from dead bacteria; turbidimetry requires a constrained linear concentration range.Source: §1.1 · pp. 3–4

1.2 Motivation of this review

• Since the 1980s, digital image processing has expanded in microorganism counting, with machine learning and deep learning accelerating after 2015 partly through more accurate segmentation of adherent colonies.The review attributes the field’s rapid growth to developments that improve image segmentation and thereby enable more precise counting.Source: §1.2 · pp. 4–5

1.3 Related reviews

• Existing reviews address selected organisms, technologies, or broader image-analysis topics, but none specifically targets image-analysis-based microorganism counting.This review therefore focuses on counting-method development and prospects, drawing on more than 136 papers for microorganism counting.Source: §1.3 · pp. 5–6

1.4 Microorganism counting methods

• The proposed counting workflow covers data acquisition, microscopy, preprocessing, segmentation and morphological operations, classification, counting, and evaluation.The review discusses thresholding, edge detection, watershed extraction, PCA, SVM, neural networks, CNNs, and metrics including accuracy, PA, MPA, and MIoU.Source: §1.4 · pp. 6, 8, 9, 10

1.5 Structure of this review

• The review selects 144 microorganism-counting papers, comprising 8 reviews, 82 bacteria-counting methods, and 54 methods for other microorganisms, and organizes them by organism group and method.Sections cover bacteria, other microorganisms, commonly used counting methods and application domains, followed by the conclusion.Source: §1.5 · p. 11

Background

• Bacteria counting is contextualized by bacteria’s role in ecosystem carbon and nitrogen cycles and its connection to human daily life.The section introduces bacteria as a crucial ecosystem component before reviewing counting methods.Source: §2 · p. 11

Motivation

• Bacteria counting addresses bacterial numbers and colony size, with importance in food safety monitoring and industrial safety detection.Manual counting is described as tedious, redundant, and subjective, motivating computer image-analysis research.Source: §2 · p. 12

Method categories

• The chapter organizes bacteria-counting methods into classic approaches, machine-learning and deep-learning approaches, and third-party tools.These categories form the chapter’s three-part structure.Source: §2 · p. 12

2.1.1 Bacteria counting method based on image enhancement

• Image-enhancement methods use contrast manipulation, filtering, and Fourier–Mellin transforms to count bacteria or colonies, with reported accuracy above 95% and differences of 1–3% from manual counts.Niyazi et al. additionally measures colony area and gray-value distribution, while Mellin-based processing provides scale and rotation invariance across twelve samples.Source: §2.1.1 · pp. 12–13

2.1.2 Bacteria counting method based on thresholding

• Thresholding methods combine filtering, color-space conversion, morphology, circle detection, and clustering procedures, achieving reported errors below 3% and accuracies up to 98%.Reported results include 2.13% average percentage error, 90% accuracy, 2.5% average relative error, 86.76±9.76% average accuracy, less than 10% mean error, and 98% accuracy.Source: §2.1.2 · pp. 13–16

2.1.3 Bacteria counting method based on edge detection

• Edge-detection pipelines enhance, smooth, threshold, fill, and filter colony boundaries, while a Gaussian–Laplacian variant achieved the best reported accuracy of 99%.The pipelines also separate clustered colonies and filter objects by size and circularity before measurement.Source: §2.1.3 · pp. 16–17

2.1.4 Bacteria counting method based on watershed

• Watershed methods separate merged colonies after thresholding or distance transforms, with classification used in some pipelines and a reported average relative error of 0.2%.Zhu et al. combine watershed segmentation, morphological opening, and Bayes classification of groups containing one to four colonies.Source: §2.1.4 · pp. 17–18

2.1.5 Bacteria counting method based on color segmentation

• Color-segmentation methods distinguish bacteria, debris, and stained cell types through multichannel processing, then calculate cell counts and sizes from area, perimeter, or local gray-value maxima.Ogawa et al. report 95% confidence intervals of the regression for simultaneous differentiation of triple-stained bacteria.Source: §2.1.5 · p. 19

PCA-based methods

• Over 99% colony segmentation and counting accuracy was achieved after PCA-based separation, local-maxima detection, and nearest-neighbor splitting of touching colonies (Yoon et al. (2015)).The local absorbance maxima search and touching-colony separation are shown in Fig. 18.Source: §2.2 · p. 19, Figure 18 · p. 20
• A PCA, distance-transform, watershed, and bottom-hat workflow achieved a mean absolute percentage error of 3.37% for colony counting.Distance transform and watershed divided overlapping colonies, while bottom-hat transformation extracted colonies from rim images.Source: §2.2 · p. 20

SVM methods

• SVM-based bacteria and colony classification reached 99.2% accuracy in chromogenic-medium analysis and relative error below 3% for rod-shaped bacteria.The methods combined background correction, segmentation, feature extraction, and supervised SVM classification (Andreini et al. (2015, 2016); Zhang et al. (2010); Chen and Zhang (2009)).Source: §2.2 · p. 20

ANN methods

• An ANN-based edge-detection pipeline analyzed 100 images per hour with 5% variation in cell counts between filters (Blackburn et al. (1998)).The ANN used six input nodes, five intermediate nodes, and three output nodes after Marr-Hildreth edge detection and filtering.Source: §2.2 · pp. 20–21, Figure 21 · p. 21

BP neural-network methods

• BP neural-network methods achieved detection error no more than 5%, while one system analyzed samples in less than 10 minutes versus 48 hours manually.Median or adaptive filtering, segmentation, morphology, and morphology-colorimetry features supported identification and counting.Source: §2.2 · pp. 21–22

Deep learning counting methods

• 92.8% accuracy was obtained with a CNN for bacterial colony classification and counting, while watershed separation yielded 92.1% accuracy and cCNN counting achieved 86%.CNN training used augmentation and preprocessing; cCNN processing was 3.8X faster with an NVIDIA Quadro K620 GPU.Source: §2.2 · p. 22

Real-time bacterial counting

• Jung and Lee (2016) reconstructed time-series high-resolution bacterial microcolony images with SPSM, segmented them, and calculated colony equivalent diameters and counts.The reconstructed images and processing workflow are shown in Fig. 25.Source: §2.3 · pp. 22–23, Figure 25 · p. 23

Commercial and automated systems

• Cellstat used quantitative RNA/DNA staining and parameterized object recognition to identify bacteria, distinguish single cells from clumps, and separately measure microcolonies and single cells.Its automated analysis was described as operator-independent and capable of analyzing many objects for improved statistics.Source: §2.3 · pp. 23–24, Figure 26 · p. 24

Agreement with reference counts

• Several systems closely matched reference counts: Model 2000 showed no statistical difference from manual counts, ProtoCOL correlated highly but slightly underestimated true counts, and KS400 achieved a correlation coefficient of 0.984.KS400 combined DAPI and Cy3 analyses with a logical AND operation to count signals from target organisms.Source: §2.3 · pp. 24–25

Classical image-processing workflows

• Classical commercial workflows used Gaussian, Laplace, gradient, thresholding, morphological, contour, and edge-effect operations to remove noise, segment cells, reject artifacts, and measure counts or biovolume.These procedures were implemented in Kontron Vidas Plus, Skidaway Tools, IBAS 2.0, Zeiss KS300, Image-Pro Plus, and Optimas systems.Source: §2.3 · pp. 24–25

Commercial and automated systems

• The Davinci method combined image enhancement, median filtering, and local binary fitting segmentation, achieving an average error of no more than 1.6% versus manual counting.Connected-region counts and areas were used to quantify bacteria.Source: §2.3 · pp. 25–26

Fluorescence and microcolony counting

• Fluorescence- and threshold-based tools supported colony and biofilm enumeration: SigmaScan Pro found no significant difference from manual counts, while ImageJ counted filled microcolonies above a selected size.MACS used automated scanning, blue LED illumination, CCD imaging, and SYBR Green II staining for microcolony counting.Source: §2.3 · pp. 25–26

Development trend

• Image-analysis approaches for bacteria counting began in the 1980s and developed rapidly during the 2010s.The review attributes this growth to increased attention to bacteria-related problems and advances in computer-based image analysis.Source: §2.4 · pp. 26–27

Counting techniques

• The most frequently used preprocessing methods are median and Gaussian filters, while thresholding, distance transform, and watershed dominate segmentation.Classifier algorithms most often reported are SVMs and ANNs.Source: §2.4 · p. 27, Table 1 · p. 29
• The reviewed literature spans diverse bacteria-counting pipelines that combine preprocessing, segmentation, classification, and evaluation methods.Table 1 organizes related work by microorganism type, preprocessing, segmentation, classification, and evaluation.Source: §2.4 · p. 26, Table 1 · p. 29
• Reported bacteria-counting results include 98% accuracy and 90% accuracy in the summarized studies.These values appear among the evaluation results compiled in Table 1.Source: Table 1 · p. 29, §2.4 · p. 31

• Bacteria-counting methods included thresholding, edge detection, sigma smoothing, median filtering, local binarization, morphological operations, contour extraction, image enhancement, and filling holes.Source: §2.4 · pp. 31–32
• A reported bacteria-counting approach achieved a 0.984 correlation and 98.4% accuracy through polynomial fitting.Source: §2.4 · p. 32
• The reviewed studies used image-analysis tools, including Image-Pro Plus, Optimas 6.5, ProtoCOL, KS400, SigmaScan Pro 5.0, and ImageJ software version 1.52.Source: §2.4 · p. 32
• Automated bacteria-counting systems included Wang et al.’s (2007) micro-colony auto counting system MACS and related software-based workflows.Source: §2.4 · p. 32

3.1.1 Counting methods based on image enhancement

• Image-enhancement methods improve microorganism separation through color filtering, contour enhancement, histogram equalization, smoothing, morphological filtering, optimized quadrats, and spatial scatter plots.Barbedo (2012b) reported more than 90% correct object identification with 8% overall deviation, while Dazzo and Gross (2013) used optimized grid sizes and centroid plots to reduce calculation complexity and edge effects.Source: §3.1.1 · p. 33

3.1.2 Counting methods based on thresholding

• Thresholding methods count fungi, mycelia, yeast, algae, protozoa, spores, and cells using global, adaptive, Otsu, HSI, histogram, and multi-threshold segmentation.Reported results include more than 94% mean algae accuracy, more than 90% algae-identification accuracy, 2.56% average relative fungi-counting error, and above 80% pus-cell reliability.Source: §3.1.2 · pp. 33–36

3.1.3 Counting methods based on edge detection

• Edge-based methods use Marr-Hildreth, Snake, Sobel, gradient, probability, Laplacian-of-Gaussian, Canny, and region-growing techniques to delineate and count microorganisms or related objects.Kocak et al. (1999) achieved 94.12% accuracy for plankton counting, while other methods measured colonies, stained cells, filaments, or animal regions against manual counts.Source: §3.1.3 · pp. 36–38

3.2 Machine learning and deep learning counting methods

• Machine-learning approaches classify counted objects using neural networks, discriminant analysis, PCA, random forests, and transfer learning after image preprocessing and feature extraction.Embleton et al. (2003) processed 75 images per sample in seven minutes, followed by 30–40 minutes of processing and classification; Benyon et al. (1999) reported 98% genus-comparison accuracy using seven basic features.Source: §3.2 · pp. 39–40

3.3 Third-party tools

• COVASIAM estimated 95.47% (σ = 8.55%) of manual colony counts versus 76% (σ = 16.27%) for single-threshold segmentation, producing 135 CFU or 97.8% of manual counts.The commercial-tool studies also reported no significant differences from manual counts for several systems and at least tenfold faster worm counting with ≤5% error.Source: §3.3 · p. 42
• Automated third-party tools achieved close agreement with manual counting, including 97.7% (± 1.0) for bacteria, 92.2% (± 2.4) for archaea, and r2>0.9 correlation for microbial cells.AnalySIS used contrast enhancement, fluorescent-probe thresholding, and pixel-based enumeration, while Automated Cell Measuring and Enumeration tool 2.0 measured object area, circularity, mean gray value, and signal-to-background ratio.Source: §3.3 · p. 44
• ImageJ and CellC workflows counted cyanobacteria after threshold-based background separation, morphological denoising, clustered-object separation, and filament-cell analysis, while watershed splitting supported yeast colony counting.The cyanobacteria analysis is shown in Fig. 43, and the yeast colony workflow is shown in Fig. 44.Source: §3.3 · p. 44, Figure 43 · p. 44, Figure 44 · p. 45

3.4 Summary of image analysis based counting for other microorganisms

• Research on other microorganism counting developed from the 1980s but remains relatively limited because algae and fungi have complex structures, while bacteria offer more abundant datasets for system evaluation.Complex hyphae make precise segmentation difficult, and systems designed for general microorganism types are often tested with bacteria datasets instead.Source: §3.4 · pp. 44–45
• The most frequently used preprocessing methods are medial and Gaussian filters, image segmentation uses thresholding and Otsu thresholding, and classifier algorithms include PCA and neural networks.Source: §3.4 · p. 45

4 Analysis of image processing based counting methods

• The review identifies preprocessing, segmentation, classification, connected-region detection, and feature extraction as the most effective components of image-based microorganism counting.These methods are analyzed according to their properties and application domains, using representative works to relate techniques to applications.Source: §4 · p. 50

4.1 Image pre-processing methods

• Preprocessing converts color images to grayscale or HSI, corrects illumination, removes noise and halos, and enhances contrast before segmentation.Examples include background subtraction, linear gray-scale transformation, low-pass, median and Gaussian filtering, morphological opening and closing, histogram equalization, CLAHE, and linear histogram expansion.Source: §4.1 · pp. 50–51

4.2 Image segmentation methods based on thresholding

• Thresholding is the basic segmentation technique: global and Otsu methods are simple, whereas adaptive local and multilevel thresholds address uneven illumination, overlapping gray levels, and lost detail.Otsu thresholding is sensitive to noise and generally follows denoising; local thresholds adapt to regional brightness distributions.Source: §4.2 · pp. 51–52, Figure 48 · p. 52, Figure 49 · p. 52

4.3 Other image segmentation methods

• Other segmentation methods include edge operators, distance-transform watershed, marker-controlled watershed, and morphological top-hat or bottom-hat transforms for connected colonies and uneven illumination.Canny performs best among the described edge operators but is relatively complex, while marker-controlled watershed reduces over-segmentation caused by noise and local discontinuity.Source: §4.3 · pp. 53–54

4.4 Image classification methods

• CNN-based classification automatically extracts image features and obtained 92.8% accuracy for bacteria colony counting in Ferrari et al. (2015), while deep-learning segmentation still requires manual adjustment.SVM achieved 99.67% bacteria-classification accuracy in Yujie (2009), and ANN classified about 95% of objects per image in Blackburn et al. (1998).Source: §4.4 · pp. 54–55

4.5 Analysis of potential methods

• Deep learning increasingly improves microorganism counting, but current methods mainly classify while segmentation still relies on thresholding or watershed; semantic segmentation is proposed as the future direction.The review also anticipates integrated systems combining sampling, imaging, and analysis for timely monitoring.Source: §4.5 · pp. 56–58

5 Conclusion and future work

• The review groups counting methods by microorganism type and segmentation approach, covering thresholding, edge detection, third-party tools, and deep learning.Classical methods developed from the 1980s to the 2000s, while deep learning advanced counting accuracy from the 2010s onward.Source: §5 · p. 58
• Professional counting systems such as ImageJ and CellC reflect increasing attention to microorganism counting and the field’s substantial research potential.The review also analyzes frequently used preprocessing, segmentation, and classification approaches.Source: §5 · p. 58
• The summarized methods can support microorganism classification, cell segmentation, industrial microscopic analysis, and small-object detection in other fields.Examples include environmental, cervical, and blood-cell classification; cancer and stem-cell segmentation; wastewater and milk monitoring; and sperm, crowd, and vehicle counting.Source: §5 · p. 58
• Deep-learning-based microorganism counting remains promising, while microorganism research retains important limitations and opportunities, especially following increased attention since COVID-19 emerged in 2019.The review is intended to contribute to future microorganism-counting research.Source: §5 · pp. 58–59