Nondestructive Assessment of Woody Breast Myopathy in Chicken Fillets Using Optical Coherence Tomography Imaging with Machine Learning

doi:10.21203/rs.3.rs-3229288/v1

Download PDF

Research Article

Nondestructive Assessment of Woody Breast Myopathy in Chicken Fillets Using Optical Coherence Tomography Imaging with Machine Learning

https://doi.org/10.21203/rs.3.rs-3229288/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Woody breast (WB) myopathy is a major muscle abnormality in chicken fillets, causing excessive hardness and chewiness. The WB condition can potentially cause big economical losses in the poultry industry by decreasing meat quality, increasing waste, degrading nutritional content, and reducing customer satisfaction. A histological technique using a light microscope has been the gold standard to characterize the sub-surface properties of the muscle with the WB condition, which is destructive, costly, time-consuming, and limited to analyzing only small sample areas. It is currently very challenging to assess the degree of WB myopathy objectively and rapidly in individual fillets. There is a need to develop an effective sensing technology for rapidly characterizing the WB condition by measuring the sub-surface cross-sections of the entire fillet at a high resolution. In this study, we utilized optical coherence tomography (OCT) to image the sub-surface microstructure of chicken muscle tissue along the entire fillet with a micrometer resolution. The OCT images provided valuable microstructural features, which were further analyzed using machine learning models to classify chicken fillets based on the WB severity. The results demonstrated a detection accuracy of up to 100% in detecting severe WB samples. The machine learning models achieved a classification accuracy of 93.3% in distinguishing normal from WB fillets. Overall, the successful application of large-scale OCT imaging demonstrated its effectiveness as a non-invasive method for evaluating WB in chicken meat. Furthermore, the study suggests that OCT imaging holds the potential for evaluating other agricultural and food products.

Poultry

Woody breast condition

Automated

Non-destructive

OCT imaging

Machine Learning

There has been a growing prevalence of woody breast (WB) myopathy in chicken breast fillets (Pectoralis major) globally which is believed to be due to the changes in the way that chickens are bred and raised for higher meat production in the poultry industry (Siddique et al., 2022). WB fillets have abnormal characteristics such as hard tissue, pale color, and hemorrhages in lesion areas (Liu et al., 2020). These symptoms can cause a decline in chicken meat quality, leading to increased waste, reduced nutritional content, decreased consumer satisfaction, and consecutive economic losses. It has been estimated that severe WB occurrence can result in an annual loss of up to $1 billion in North America (Che et al., 2022). Therefore, several attempts have been made to characterize and sort WB chicken fillets (Chatterjee et al., 2016; Geronimo et al., 2019; Kuttappan et al., 2016; Petracci et al., 2019; Tijare et al., 2016).

Typically, standard light microscopy and transmission electron microscopy methods are used for the evaluation of the quality and safety of food products, including assessing chicken meat for WB disorder by characterizing its microstructure (Li et al., 2021). However, these methods are associated with certain drawbacks, such as being destructive, costly, time-consuming, and restricted to analyzing only small sample areas. Additionally, the distribution of WB areas on chicken muscle is heterogynous, with varying degrees of WB areas present along the chicken fillet in moderate and severe cases (Soglia et al., 2017; Tasoniero et al., 2017; Wold et al., 2019). As a result, evaluating the level of WB condition in the entire chicken fillet only based on a small scan area can be misleading. Therefore, there is a need for the development of effective, rapid, and large-scale noninvasive techniques for characterizing the WB condition by measuring the sub-surface cross-sections of the entire fillet at a high resolution.

There are various nondestructive techniques for the evaluation of food quality and safety such as spectroscopy, electronic nose/tongue, hyperspectral imaging (HSI), ultrasound, magnetic resonance imaging (MRI), optical coherence tomography (OCT), confocal microscopy, X-ray, and computed tomography (CT) (Adedeji et al., 2020; Shi et al., 2021). One major limitation of most non-destructive methods is their inability to provide high-resolution spatial and depth information simultaneously. Although the HSI method captures both spectral and spatial information, it does not provide depth visualization (Zhou et al., 2022). Conversely, while ultrasound, X-ray imaging, and computed tomography are depth imaging techniques, the standard devices based on these technologies often provide low lateral and depth resolution at the millimeter scale (Byun et al., 2019). Confocal microscopy can generate high-resolution internal 3D images, but its depth of penetration is much shallower than that of the other high-resolution depth imaging technique, namely the OCT method.

OCT is a contactless cross-sectional imaging technique that utilizes low-coherence light to produce high-resolution 2D and 3D structural images of translucent and even opaque materials by capturing and processing the interfered light reflected from the internal tissues (Podoleanu, 2012). In OCT, near-infrared light at approximately 800–1300 nm is commonly used for analyzing biological materials as it has minimal absorption by water and maximum penetration deep underneath the sample skin (Fujimoto & Drexler, 2015). OCT has a distinct advantage over other methods as it can capture both the spatial as well as depth (lateral and axial) information with micrometer resolution up to a few millimeters deep under the surface of the food sample (Li et al., 2019). Furthermore, OCT offers a notable advantage in terms of speed compared to other imaging methods. It can provide approximately 29,000 axial A-scans per second, and in ultrahigh-speed OCT devices, this rate can even exceed 300,00 scans per second (Li et al., 2021). OCT can also provide cross-sectional B-scan images at a rate of more than 100 frames per second. Therefore, this study hypothesizes that the choice of OCT modality enables the development of an imaging system for the characterization and assessment of WB in chicken fillets through rapidly capturing high-resolution sub-surface images of tissues in a non-invasive and safe manner, with broad spatial coverage along the fillet.

In recent years, several studies have explored the application of OCT in agriculture and food-related fields, such as detecting disease in seeds and kernels (Lee et al., 2011), analyzing dried fruit tissue (Rizzolo et al., 2014), investigating the structural and tissue properties of fruits and vegetables (Li et al., 2015; Verboven et al., 2013), and sub-skin bruise detection in fruits (Zhou et al., 2018, 2021). Zhou et al. (2022) used OCT for automatic non-invasive early detection of mold on maize kernels based on locating the mold attached to the upper pericarp layer. They developed an algorithm to automatically locate and segment the mold area employing denoising, edge detection, smoothing, flattening, cropping, segmentation, object detection, and decision-making steps. While they achieved promising results for moldy maize seed detection, they assessed only a single small field of view (FOV) within a sample with OCT imaging, which could potentially miss the infected areas in other locations. Korzeniewska et al. (2020) studied the feasibility of using OCT imaging for detecting the change in the tissue structure of raspberries and grapes subjected to a pulsed electric field (PEF). They observed significant tissue texture changes in the 2D B-scans of both fruits after PEF treatment by comparing the quantities of textural features of the images such as entropy, mean, standard deviation, and histogram. They used one drupelet of each fruit in the PEF experiments and subsequent OCT imaging. A similar study was conducted by Goclawski et al. (2017) to investigate the feasibility of using OCT imaging for detecting the sub-skin structural changes in grapes after PEF treatment. While they observed some differences in the value of the textural features extracted from the OCT images for the control and PEF-treated samples, they did not provide statistical or machinelearning tests on a larger dataset to examine the accuracy of the technique in classifying the samples. In another study, OCT imaging was used for the characterization of the mechanical and structural properties of blueberries during storage (Li et al., 2021). They acquired 3D OCT images of each blueberry and calculated the thickness of the two top near-surface layers of the cells in the epidermal and subepidermal areas of the fruit as their features. Their results showed a significant increase in the thickness of the top layer with an increase in water loss in blueberries. Although the results of these studies have revealed great potential for providing accurate micro-resolution depth information from agricultural products, there are some gaps in the use of OCT in food quality evaluation, such as the need for developing automated algorithms to interpret and extract useful information from OCT images. In addition, the data collection time required to get comprehensive information, particularly from larger samples, is quite high. Hence, further research is necessary to develop imaging systems based on OCT for robust quality evaluation analysis with faster and more accurate results.

OCT imaging with a B-scan mode provides a relatively small FOV compared to other imaging methods with a maximum lateral dimension of around 10 mm. It means that relying solely on one OCT B-scan image may not provide comprehensive information for large and heterogeneous samples. There is no study found in the literature that used OCT for large-scale scanning of chicken meat. Only in one study, Yoon et al. (2016) conducted a preliminary investigation into the viability of utilizing OCT for quantifying the sub-surface structure of WB chicken fillets. However, their study had limitations, as they only analyzed a small section of a fillet (56 mm long) and evaluated a limited number of samples, without conducting statistical or machine learning analysis on the extracted features. In this study, OCT was used as a subsurface cross-sectional imaging approach in the form of large-scale scanning along a whole fillet using a semi-automated imaging system for classifying fillets with different levels of WB severity. The acquired images for each fillet were stitched together to be analyzed using an automated image processing algorithm to extract relevant features to be used to train machine learning models to classify or predict WB characteristics of the fillet. Hence, the main objective of this study was to predict and assess the severity of WB in fillets using OCT cross-sectional imaging, image processing, and machine learning techniques, and to classify fillets into two or three classes based on the WB condition. The sub-objectives were: (1) to determine the feasibility of detecting WB using nondestructive OCT imaging of tissue section, (2) To extract relevant features from OCT images that exhibit a high correlation with the WB condition (3) to study and analyze the spatial distribution of epimysium thickness and WB areas across an entire fillet.

Chicken Fillets

The chicken fillets (pectoralis major) used in this study were collected from the deboning line of a commercial plant approximately 3 hr. after the slaughter. They were transported to the U.S. National Poultry Research Center in Athens, GA in ice-covered bags within approximately 45 min of collection. The samples were trimmed to remove excess fat and connective tissue, although some small traces of fat remained after the trimming process. An expert panel scored the samples as normal, moderate, and severe based on the WB severity level on the same experiment day before doing the OCT image acquisition. A total of 30 chicken fillets were examined over two consecutive experiment days, with 12 samples collected on the first experimental day and 18 samples (refrigerated at 5°C overnight) collected on the second day. This study utilized a total of 10 fillets from each of the three classes. Figure 1 shows photos of typical fillets from each of the three classes, showcasing three distinct levels of WB condition.

OCT Device and Data Acquisition

The imaging system used in this study was a spectral-domain optical coherence tomography (SD-OCT) (Callisto, Thorlabs, Newton, NJ, USA) consisting of four main components: a base unit, a personal computer (PC), a probe head, and a probe stand (Fig. 2). The base unit acted as the control center for the SD-OCT system, while the PC was used to manage the imaging process and process the data obtained. The base unit housed a 930-nm broadband super luminescent diode (SLD) with a spectral bandwidth of 100 nm as the light source. The light travels through a fiber optic coupler to the probe head, where it is split into a reference arm and a sample arm. The probe head has an interferometer unit where the back-scattered light from the sample is recombined with the back-reflected reference light from the fixed reference mirror. The interference light is sent back to the base unit where the spectrometer analyzes the phase delay at each wavelength by dispersing the light spectrally and detecting it using a linear CCD array. The Fast Fourier Transform (FFT) was used to obtain the depth profile of the sample from spectral data. To capture a cross-sectional image, the light beam was scanned over a 5 mm range of the sample surface using a galvanometer mirror located in the probe head. During the scanning process, a total of 2028 interferograms were recorded. These interferograms were then analyzed using FFT to create 1D A-scans, which were further reconstructed into 2D images, also known as B-scans. Thus, each B-scan consisted of 2028 A-scans. The FOV for each B-scan was 5 mm (X, lateral) by 1.7 mm (Z, depth) with a lateral and depth resolution of 2.44 and 3.31 µm, respectively. The time required to acquire each B-scan was 1.71 seconds.

To obtain a comprehensive view of each fillet sample, its central line was scanned along its three major portions, namely the cranial, middle, and caudal parts. This was accomplished by acquiring a total of 25 B-scan images from each fillet, with the sample being moved under the OCT probe using a motorized stage. After acquiring one B-scan, the stage was moved by 5 mm relative to the previous position to the next starting point, thus ensuring complete coverage of the targeted area. In this study, the movement of the motorized stage and acquisition of OCT images were manually controlled using the application software provided by the manufacturers. Also, the working distance of the OCT probe from the sample surface was manually adjusted at each position before acquiring each B-scan image. A total of 750 B-scan images were obtained in this study to be used in the image processing and machine learning sections.

OCT Data Processing and Analysis

The OCT cross-sectional images of normal and severe WB chicken fillets are shown in Fig. 3. Since the preliminary analysis of OCT data showed a difference in the microstructure of the top layer of the chicken breast muscle with and without WB condition, the objective of the proposed method was to automatically extract and quantify the upper layer of the chicken fillet, known as the epimysium, to assess the severity of WB in fillets.

Figure 4 shows the flowchart outlining the steps involved in OCT data acquisition and analysis. Initially, the B-scans were converted from raw 24-bit PNG format to grayscale images. To ensure consistency, the image edges were cropped to remove pixels with zero values. Subsequently, a median filter with a window size of 10×10 was applied to suppress salt and pepper noise as well as speckle noise present in the images. The upper section of all B-scans, representing the air above the sample, displayed a uniformly low-intensity image section. To eliminate the non-informative background and isolate the target layers, two distinct thresholding approaches were implemented. These methods aimed to differentiate the desired layers from the remaining image content based on their intensity values. In the first approach, a predefined threshold value derived from the intensity histogram of the images was utilized. In the second approach, the Otsu multi-thresholding method was employed to segment the two top layers of the chicken fillet in each OCT image. This resulted in an image consisting of three distinct intensity levels, representing the top layers and the background, as illustrated in Fig. 5. The use of a multi-thresholding approach was beneficial as it allowed for cleaner segmentation of the top layer without interfering with the underlying layer, enabling the separation of the under layer as a separate feature. Following the segmentation, the thickness of each layer was measured perpendicular to the sample surface to obtain accurate measurements. This required flattening of the image to ensure a horizontal surface. The flattening process involved identifying the boundary of the top layer by locating the first maximum value in each column and fitting a line to these boundary points. Subsequently, the image was rotated to align the fitted line horizontally, resulting in a flattened image. To address the issue of decreased intensity in the sloped areas, caused by the higher distance from the OCT probe, which led to segmented layers with lower thickness values than the upper side of the image, a solution was implemented. This involved automatically cropping either the left or right half of the image, depending on whether it fell under the central line in images with a sharp negative or positive slope value in the fitted boundary line, respectively. After the images were segmented into three distinct regions and flattened, the separate images of the fillet were stitched to produce a consolidated image for each sample (Fig. 6). This was carried out to account for the heterogeneous distribution of WB along the fillets. Instead of extracting features from each individual image and labeling each image based on the fillet sample, this approach extracted the features fillet-wise. Subsequently, the extracted features for each fillet were utilized to develop classification models to classify each fillet.

All OCT image processing and data analysis were conducted using Matlab version R2022b (The MathWorks Inc., Natick, Massachusetts, USA).

Feature Extraction

This study involved the extraction of certain morphological and textural characteristics from the fillet samples, which were then used as input variables for the machine learning models. The first features were the average and standard deviation of the thickness of the uppermost layer, namely the epimysium layer, along each fillet. Calculating the thickness of the top layer at each point on the sample surface was carried out by counting the number of non-zero pixels along each column (A-scan). This measurement was conducted using the thickness values obtained from the two different thresholding approaches previously discussed, giving two sets of thickness features. The next feature was based on the proportion of the thickness of the uppermost layer to that of the second uppermost layer. To calculate a fillet-wise value for this feature, first, a pre-defined threshold was established. Then, the number of crossings of the feature value in each fillet with respect to the threshold was counted to identify WB areas. Next, the percentage of WBness in each fillet was calculated by dividing the WB area by the total fillet area. This provided a measure of the proportion of each fillet that is WB.

Machine Learning Classification Models

To classify normal and WB chicken fillets based on the extracted features, machine learning procedures were developed using various classification algorithms, including support vector machine (SVM), random forest (RF), k-nearest neighbors (kNN), linear discriminant analysis (LDA), decision tree (DT), and Extreme Gradient Boosting (XGB). The classifiers were evaluated based on metrics such as accuracy, recall, precision, F1-score, and kappa coefficient. In this study, the repeated holdout evaluation method was employed to assess the performance of the model which involves multiple iterations of dataset partitioning into training and testing sets. The dataset was divided into 80% for training and 20% for testing, and this partitioning process was repeated five times. During each iteration, the model was trained on the training data and evaluated using the corresponding testing data. The performance metric was then calculated by taking the average testing accuracy across these five repetitions. This will give a more reliable estimation of the model's accuracy compared to a single partitioning of the dataset. Taking the average accuracy over multiple repetitions helps to mitigate the impact of random fluctuations and provides a more stable evaluation of the model's generalization performance. All machine learning processes utilized in this study for the classification of chicken fillet samples were conducted on the Python 3.10 platform (Python Software Foundation, https://www.python.org) and implemented in the PyCharm Editor. The open-source libraries of Pandas, Numpy, Scikit-learning, and Matplotlib were used in this work.

Visualization and Observations

In Fig. 7, mosaics of surface color video frames are presented, where OCT B-scan lines are overlaid for a typical normal fillet and two WB fillets. Each video frame contains a red horizontal arrow denoting the scan line position. Furthermore, below the video mosaic, Fig. 7 includes mosaics of corresponding B-scan images obtained from the same scan line. In Fig. 7, the visual analysis reveals that the color distinction between the normal and WB fillet images is relatively small. However, specific lesion areas can be observed, particularly reddish spots that exclusively correspond to petechial hemorrhagic lesions found in the severe WB sample in Fig. 7 (c). These findings are consistent with previous research, which has documented lesion areas particularly hemorrhagic lesions in WB samples (Kuttappan et al., 2017; Soglia et al., 2017). On the other hand, as shown in Fig. 7 (b), not all WB samples have those lesion areas. Furthermore, both the normal and WB samples can exhibit fatty areas and white stripes visible on their surfaces. This complex visual presentation underscores the challenges associated with relying solely on surface color images to detect WB. The overlapping features and similarities between normal and WB samples pose difficulties in visually identifying WB based on surface appearance alone.

However, examining the OCT B-scans and segmented depth images provides valuable insights. It is evident that the thickness of the epimysium layer, which is the top layer on the surface of the sample, is significantly larger in the WB sample images compared to the normal sample. These findings are consistent with the previous study conducted by Yoon et al. (2016), which reported that the epimysium thickness of wooden breast fillets in OCT images was approximately twice as large as that of normal fillets. This is also in agreement with the finding of previous studies suggesting that the WB condition results in the accumulation of collagen and other connective tissues in the breast muscle, leading to the thickening of the connective tissue (Soglia et al., 2021). Therefore, the observation in this study emphasizes the potential of OCT imaging and segmented depth analysis for distinguishing WB fillets based on the increased thickness of the epimysium layer. These structural differences offer a promising method for WB detection beyond surface color images.

In Fig. 8, four examples of video frames, OCT scan lines, B scans, and A scans are shown for facilitating a comparison between a normal fillet and three severe WB fillets. Notably, Fig. 8 highlights variations observed in the severe WB fillets, such as the presence or absence of white stripes and lesion areas. However, it is crucial to emphasize that regardless of these variations, the thickness of the epimysium layer in WB samples consistently exceeds that of the normal sample. This distinction becomes particularly noticeable in the WB samples exhibiting lesion areas and prominent white stripes.

Histogram Analysis

To further analyze the difference in the epimysium thickness for normal, moderate, and severe WB fillets, a histogram plot is presented in Fig. 9 illustrating the distribution of epimysium thickness along the scanning line of the fillets. The x-axis represents thickness values, divided into equally sized bins, while the y-axis represents the normalized probability distribution function of points within each bin. From the histogram plot in Fig. 9, several noteworthy observations can be made. Firstly, all three groups exhibit a bell-shaped distribution, but the normal group shows a slight left-skewness and a narrower shape. This suggests that a significant portion of normal samples had thickness values toward the lower end of the range, with only a minimal percentage of areas having high epimysium thickness values. Secondly, although there is some overlap between normal and severe WB samples, indicating a non-uniform distribution of WB areas, a clear distinction can be observed between these two sample groups. However, the moderate samples fall in between the other two groups, posing a challenge in separating them from the normal and severe groups. Moreover, the distribution of WB samples appears symmetrical with a wider shape, with a peak around 90 for thickness values. These observations indicate that it might be possible to classify the three fillet groups based on the thickness of the epimysium layer extracted from the OCT images. However, the moderate group's placement between the normal and severe groups suggests a need for further investigation to establish a more definitive classification criterion.

Woody Breast Condition Classification

Table 1 provides an overview of the classification performance on both the training and testing datasets achieved by various classifiers on different classification tasks related to the WB condition. For the classification of normal and severe WB fillets, all classifiers demonstrated relatively high accuracy. XGB classifier achieved perfect accuracy of 100% on the training and test dataset, while DT and RF classifiers achieved high test accuracies of 96.0% and 92.0%, respectively. These results indicate that the classifiers were able to effectively differentiate between normal and severe WB fillets. For the classification of normal and WB (moderate and severe combined), the classifiers also achieved high accuracies. XGB, DL, and NB performed well with an accuracy of 93.3% on the testing dataset, followed closely by KNN with an accuracy of 90.0%. These results suggest that the classifiers were reasonably successful in identifying any kinds of WB samples from the normal ones. It is important to emphasize that the task of sorting WB samples from normal ones holds significant practical relevance in real-world industry applications. By achieving high accuracies in this task, the classifiers showcased their potential for reliable identification and differentiation of WB samples, providing valuable insights for quality assessment and control in the industry. Moreover, the highest classification accuracy achieved for distinguishing between normal and moderate WB fillets was 95.0% using the DT model. In contrast, the maximum accuracy rate for differentiating moderate WB from severe WB fillets was 85.0%. These results highlight the difficulty of accurately separating moderate WB samples from severe WB samples, making it the most challenging task among the classification tasks analyzed. However, the results of this study showed a significant improvement in classifying moderate and severe WB fillets compared to the results of a similar work by Wold et al. (2019) who obtained 76.9% training and 68.7% test accuracy for the same task using NIR spectroscopy. Lastly, for the three-class classification task involving normal, moderate, and severe WB fillets, the classifiers achieved varying levels of accuracy. Three-class modeling resulted in noticeably lower accuracies than that in all the two-class models. Its highest rate was achieved using RF with an accuracy of 80.0%, followed by NB with an accuracy of 73.3%. This can be as a result of adding the moderate class which is harder to be separated from the severe class. Overall, the high accuracy rates achieved on various tasks indicate the feasibility of using OCT imaging combined with machine learning for automated detection and classification of the WB condition in poultry meat production. This improvement in the discrimination of normal and WB fillets in this study compared to previous studies can be explained by two main considerations. First, the large-scale scanning across the fillet surface provided more representative information from the spatially unevenly distributed WB areas. And secondly the micron level measurement of the properties of the epimysium layer, which comprises collagen molecules and contributes to the hardness of the WB fillets, performed as a more relevant feature for the WB detection. However, further refinement and optimization may be necessary, particularly in distinguishing between moderate and severe WB samples, by obtaining information from a larger area of individual fillets and fusing other highly relevant features.

Table 1

Performance of different classifiers for the classification of normal vs. woody breast chicken fillets using OCT imaging
Classifier	Normal breast vs. Severe WB		Normal breast vs. WB		Normal breast vs. Moderate WB		Moderate WB vs. Severe WB		3 Classes
Classifier	Train	Test	Train	Test	Train	Test	Train	Test	Train	Test
XGB	100	100	91.3	93.3	89.2	85.0	85.0	85.0	95.6	60.0
SVM	92.9	84.0	92.1	86.6	98.4	90.0	80.0	60.0	76.1	70.0
DT	100	96.0	100	93.3	100	95.0	100	65.0	100	66.6
RF	100	92.0	100	86.6	100	80.0	100	55.0	100	80.0
KNN	95.0	76.0	95.6	90.0	93.8	80.0	85.0	55.0	83.4	66.6
NB	84.7	68.0	86.0	93.3	92.3	85.0	85.0	85.0	81.7	73.3
LDA	83.0	80.0	86.0	80.0	63.0	50.0	80.0	75.0	61.7	46.6

XGB extreme gradient boosting, SVM support vector machine, DT decision trees, RF random forest, kNN k-nearest neighbors, NB Naïve Bayes, LDA linear discriminant analysis

Tables 2 ‒ 4 display the recall, precision, F1-score, and Kappa coefficient values obtained by various classifiers for different classification tasks related to the WB condition. In Table 2, the XGB classifier achieved the perfect scores of 1.0 for recall, precision, F1-score, and Kappa coefficient, indicating excellent performance in correctly identifying both normal and severe WB samples. The DT and RF models also performed well with score values near 1.0, indicating strong performance in correctly classifying normal and severe WB fillets. Furthermore, the XGB classifier demonstrated superior performance in classifying normal and WB samples, achieving 0.92, 0.94, 0.92, and 0.84 for recall, precision, F1-score, and Kappa coefficient scores, respectively. These metrics highlight the effectiveness of the method in accurately identifying and distinguishing between normal and WB samples. The high recall value indicates its ability to correctly classify a significant proportion of actual positive instances. The precision score reflects its capability to minimize false positives which can lead to less waste of normal samples labeled as defective. F1-score represents the harmonic mean of precision and recall, providing a balanced measure of its overall performance. The Kappa coefficient, which assesses agreement beyond chance, indicates a substantial level of agreement in the classification results and is considered to be good if it is higher than 0.8.

Table 2

Classification performance for normal vs. woody breast chicken fillets using OCT imaging
Classifier	Normal breast vs. Severe WB
Classifier	Recall	Precision	F1-score	Kappa	Recall		Precision		F1-score		Kappa
XGB	1.0	1.0	1.0	1.0		0.92		0.94		0.92		0.84
SVM	0.87	0.86	0.82	0.84		0.77		0.81		0.81		0.73
DT	0.96	0.96	0.96	0.87		0.92		0.94		0.92		0.75
RF	0.95	0.9	0.9	0.85		0.85		0.78		0.80		0.72
KNN	0.78	0.82	0.74	0.82		0.89		0.92		0.88		0.73
NB	0.66	0.62	0.63	0.71		0.92		0.94		0.92		0.75
LDA	0.81	0.72	0.73	0.69		0.78		0.79		0.74		0.73

XGB extreme gradient boosting, SVM support vector machine, DT decision trees, RF random forest, kNN k-nearest neighbors, NB Naïve Bayes, LDA linear discriminant analysis

Table 3 shows the classification performance of the models in distinguishing moderate WB samples from normal and severe WB ones. For example, in the task of classifying moderate WB vs. normal samples, the DT classifier demonstrated a recall of 0.96, precision of 0.95, F1-score of 0.94, and Kappa coefficient of 0.76. These scores were higher than the ones for classifying moderate vs. severe WB samples with a maximum of 0.90, 0.88, 0.84, and 0.74 for recall, precision, F1-score, and Kappa coefficient, respectively. This indicates that distinguishing between moderate WB fillets and severe WB fillets is more challenging compared to classifying moderate WB fillets from normal ones.

Table 3

Classification performance for normal and severe woody breast vs. moderate woody breast chicken fillets using OCT imaging
Classifier	Normal breast vs. Moderate WB
Classifier	Recall	Precision	F1-score	Kappa	Recall	Precision	F1-score	Kappa
XGB	0.81	0.79	0.77	0.6	0.9	0.88	0.84	0.74
SVM	0.91	0.91	0.89	0.7	0.71	0.56	0.57	0.57
DT	0.96	0.95	0.94	0.76	0.63	0.68	0.63	0.48
RF	0.81	0.85	0.78	0.72	0.57	0.54	0.45	0.39
KNN	0.81	0.76	0.84	0.7	0.53	0.54	0.48	0.33
NB	0.88	0.86	0.84	0.7	0.72	0.74	0.71	0.36
LDA	0.46	0.26	0.32	0.59	0.76	0.78	0.74	0.38

XGB extreme gradient boosting, SVM support vector machine, DT decision trees, RF random forest, kNN k-nearest neighbors, NB Naïve Bayes, LDA linear discriminant analysis

Table 4 presents the performance of different classifiers in the three-class classification of normal, moderate, and severe WB fillets. The RF classifier achieved relatively higher scores across the metrics, indicating better overall performance in accurately identifying the three classes. However, there is still room for improvement in achieving higher accuracy in the 3-class task.

Table 4

Classification performance for normal and severe woody breast vs. moderate woody breast chicken fillets using OCT imaging
Classifier	3-Class Classification of Normal breast, Moderate WB, and Severe WB
Classifier	Recall	Precision	F1-score	Kappa
XGB	0.62	0.62	0.56	0.38
SVM	0.70	0.69	0.64	0.43
DT	0.47	0.53	0.47	0.43
RF	0.72	0.65	0.67	0.47
KNN	0.66	0.55	0.56	0.47
NB	0.68	0.66	0.64	0.49
LDA	0.42	0.31	0.33	0.45

XGB extreme gradient boosting, SVM support vector machine, DT decision trees, RF random forest, kNN k-nearest neighbors, NB Naïve Bayes, LDA linear discriminant analysis

The findings of this study, which employed OCT imaging coupled with traditional image processing and machine learning techniques, yielded promising results in the accurate detection and classification of WB fillets. This indicates that the applied algorithms were effective in extracting relevant features and making accurate classifications. In addition, the methods used had low computational complexity without the need for large amounts of labeled data for training. However, it is important to note that these techniques relied on manual thresholding and feature extraction which may limit their adaptability to complex and diverse datasets. On the other hand, deep learning methods can achieve superior performance with automatically learning discriminative features and patterns directly from the data. Thus, in the future direction, a fully automated high-speed OCT scanning system will be developed to enable the acquisition of a large volume of data from a larger number of samples which could be utilized to enhance the performance of deep learning models. Additionally, such an automated high-speed scanning system would allow for OCT image data to be acquired from a larger portion of the fillet, rather than scanning only one central line along each fillet. This would enable a more accurate assessment of the non-uniform distribution of WB areas within each sample. This improved dataset would facilitate more accurate and reliable classification, even in the challenging three-class task involving normal, moderate, and severe WB samples. This can contribute to better understanding and management of this condition.

The application of spectral domain OCT imaging, along with image processing and machine learning methods, was investigated for the rapid and non-invasive detection and classification of WB chicken fillets. The results of this study demonstrated the high effectiveness of the combination of OCT imaging, image processing, and machine learning algorithms in distinguishing between normal and WB fillets with a maximum accuracy of 100%, 93.3%, 95%, and 85% for the normal vs. severe WB, normal vs. WB, normal vs. moderate WB, and moderate vs. severe WB cases, respectively. This novel approach showcased the potential of OCT as a rapid and non-invasive sub-surface imaging tool, capable of segmenting and quantifying the top layers of skinless chicken breast throughout the entire fillet. This has substantial implications for the poultry industry, enabling more efficient quality control measures and improved product selection. This study lays the foundation for further advancements in utilizing OCT as a valuable tool in poultry inspection and quality assessment. Further research is needed to improve the rates for the three-class classification including normal, moderate, and severe WB samples. As a future work, the development of a fully automated high-speed OCT scanning system will be investigated to facilitate the acquisition of large volumes of data for deep learning-based classification models.

Author Contribution Nader Ekramirad: data curation, formal analysis, and writing, original draft preparation, investigation, and visualization. Seung-Chul Yoon: investigation, editing, conceptualization, supervision, and funding acquisition. Brian C. Bowker: Methodology. Hong Zhuang: Methodology.

Funding This study was supported by the U.S. Department of Agriculture – Agricultural Research Service.

Data Availability The datasets obtained during the current study are available from the corresponding author upon reasonable request.

Conflict of Interest The authors declare no competing interests.

Adedeji, A. A., Ekramirad, N., Rady, A., Hamidisepehr, A., Donohue, K. D., Villanueva, R. T., Parrish, C. A., & Li, M. (2020). Non-destructive technologies for detecting insect infestation in fruits and vegetables under postharvest conditions: A critical review. Foods, 9(7), 927. https://doi.org/10.3390/foods9070927.
Byun, K. H., Cho, M. J., Park, S. Y., Chun, H. S., & Ha, S. D. (2019). Effects of gamma ray, electron beam, and X-ray on the reduction of Aspergillus flavus on red pepper powder (Capsicum annuum L.) and gochujang (red pepper paste). Food Science and Technology International, 25(8), 649–658. https://doi.org/10.1177/1082013219857019.
Chatterjee, D., Zhuang, H., Bowker, B. C., Sanchez-Brambila, G., & Rincon, A. M. (2016). Instrumental texture characteristics of broiler pectoralis major with the wooden breast condition. Poultry Science, 95(10), 2449–2454. https://doi.org/10.3382/ps/pew204.
Che, S., Wang, C., Iverson, M., Varga, C., Barbut, S., Bienzle, D., & Susta, L. (2022). Characteristics of broiler chicken breast myopathies (spaghetti meat, woody breast, white striping) in Ontario, Canada. Poultry Science, 101(4), 101747. https://doi.org/10.1016/j.psj.2022.101747.
Fujimoto, J. G., & Drexler, W. (2015). Introduction to OCT. Optical Coherence Tomography. https://doi.org/10.1007/978-3-319-06419-2_1. 3.
Geronimo, B. C., Mastelini, S. M., Carvalho, R. H., Barbon Júnior, S., Barbin, D. F., Shimokomaki, M., & Ida, E. I. (2019). Computer vision system and near-infrared spectroscopy for identification and classification of chicken with wooden breast, and physicochemical and technological characterization. Infrared Physics & Technology, 96, 303–310. https://doi.org/10.1016/j.infrared.2018.11.036.
Gocławski, J., Sekulska-Nalewajko, J., Korzeniewska, E., & Piekarska, A. (2017). The use of optical coherence tomography for the evaluation of textural changes of grapes exposed to pulsed electric field. Computers and Electronics in Agriculture, 142, 29–40. https://doi.org/10.1016/j.compag.2017.08.008.
Korzeniewska, E., Sekulska-Nalewajko, J., Gocławski, J., Dróżdż, T., & Kiełbasa, P. (2020). Analysis of changes in fruit tissue after the pulsed electric field treatment using optical coherence tomography. The European Physical Journal Applied Physics, 91(3), Article 3. https://doi.org/10.1051/epjap/2020200021.
Kuttappan, V. A., Hargis, B. M., & Owens, C. M. (2016). White striping and woody breast myopathies in the modern poultry industry: A review. Poultry Science, 95(11), 2724–2733. https://doi.org/10.3382/ps/pew216.
Kuttappan, V. A., Owens, C. M., Coon, C., Hargis, B. M., & Vazquez-Añon, M. (2017). Incidence of broiler breast myopathies at 2 different ages and its impact on selected raw meat quality parameters. Poultry Science, 96(8), 3005–3009. https://doi.org/10.3382/ps/pex072.
Lee, C., Lee, S. Y., Kim, J. Y., Jung, H. Y., & Kim, J. (2011). Optical sensing method for screening disease in melon seeds by using optical coherence tomography. Sensors (Basel, Switzerland), 11(10), Article 10. https://doi.org/10.3390/s111009467.
Li, M., Landahl, S., East, A. R., Verboven, P., & Terry, L. A. (2019). Optical coherence tomography—A review of the opportunities and challenges for postharvest quality evaluation. Postharvest Biology and Technology, 150, 9–18. https://doi.org/10.1016/j.postharvbio.2018.12.005.
Li, M., Rivera, S., Franklin, D., Nowak, E., Hallett, I., Kolenderska, S., Urbańska, M., Vanholsbeeck, F., & East, A. (2021). Use of optical coherence tomography and light microscopy for characterisation of mechanical properties and cellular level responses of ‘Centurion’ blueberries during weight loss. Journal of Food Engineering, 303, 110596. https://doi.org/10.1016/j.jfoodeng.2021.110596.
Li, M., Verboven, P., Buchsbaum, A., Cantre, D., Nicolaï, B., Heyes, J., Mowat, A., & East, A. (2015). Characterising kiwifruit (Actinidia sp.) near skin cellular structures using optical coherence tomography. Postharvest Biology and Technology, 110, 247–256. https://doi.org/10.1016/j.postharvbio.2015.08.021.
Liu, J., Puolanne, E., Schwartzkopf, M., & Arner, A. (2020). Altered sarcomeric structure and function in woody breast myopathy of avian pectoralis major muscle. Frontiers in Physiology, 11. https://www.frontiersin.org/articles/10.3389/fphys.2020.00287.
Petracci, M., Soglia, F., Madruga, M., Carvalho, L., Ida, E., & Estévez, M. (2019). Wooden-breast, white striping, and spaghetti meat: Causes, consequences and consumer perception of emerging broiler meat abnormalities. Comprehensive Reviews in Food Science and Food Safety, 18(2), 565–583. https://doi.org/10.1111/1541-4337.12431.
Podoleanu, A. G. (2012). Optical coherence tomography. Journal of microscopy, 247(3), 209–219. https://doi.org/10.1111/j.1365-2818.2012.03619.x.
Rizzolo, A., Vanoli, M., Cortellino, G., Spinelli, L., Contini, D., Herremans, E., Bongaers, E., Nemeth, A., Leitner, M., Verboven, P., Nicolaï, B. M., & Torricelli, A. (2014). Characterizing the tissue of apple air-dried and osmo-air-dried rings by X-CT and OCT and relationship with ring crispness and fruit maturity at harvest measured by TRS. Innovative Food Science & Emerging Technologies, 24, 121–130. https://doi.org/10.1016/j.ifset.2013.09.001.
Shi, Y., Wang, X., Borhan, M. S., Young, J., Newman, D., Berg, E., & Sun, X. (2021). A review on meat quality evaluation methods based on non-destructive computer vision and artificial intelligence technologies. Food Science of Animal Resources, 41(4), 563. https://doi.org/10.5851/kosfa.2021.e25.
Siddique, A., Herron, C. B., Valenta, J., Garner, L. J., Gupta, A., Sawyer, J. T., & Morey, A. (2022). Classification and feature extraction using supervised and unsupervised machine learning approach for broiler woody breast myopathy detection. Foods, 11(20), https://doi.org/10.3390/foods11203270. Article 20.
Soglia, F., Gao, J., Mazzoni, M., Puolanne, E., Cavani, C., Petracci, M., & Ertbjerg, P. (2017). Superficial and deep changes of histology, texture and particle size distribution in broiler wooden breast muscle during refrigerated storage. Poultry Science, 96(9), 3465–3472. https://doi.org/10.3382/ps/pex115.
Soglia, F., Petracci, M., Davoli, R., & Zappaterra, M. (2021). A critical review of the mechanisms involved in the occurrence of growth-related abnormalities affecting broiler chicken breast muscles. Poultry Science, 100(6), 101180. https://doi.org/10.1016/j.psj.2021.101180.
Tasoniero, G., Bertram, H. C., Young, J. F., Zotte, D., A., & Puolanne, E. (2017). Relationship between hardness and myowater properties in Wooden Breast affected chicken meat: A nuclear magnetic resonance study. LWT, 86, 20–24. https://doi.org/10.1016/j.lwt.2017.07.032.
Tijare, V. V., Yang, F. L., Kuttappan, V. A., Alvarado, C. Z., Coon, C. N., & Owens, C. M. (2016). Meat quality of broiler breast fillets with white striping and woody breast muscle myopathies. Poultry Science, 95(9), 2167–2173. https://doi.org/10.3382/ps/pew129.
Verboven, P., Nemeth, A., Abera, M. K., Bongaers, E., Daelemans, D., Estrade, P., Herremans, E., Hertog, M., Saeys, W., Vanstreels, E., Verlinden, B., Leitner, M., & Nicolaï, B. (2013). Optical coherence tomography visualizes microstructure of apple peel. Postharvest Biology and Technology, 78, 123–132. https://doi.org/10.1016/j.postharvbio.2012.12.020.
Wold, J. P., Måge, I., Løvland, A., Sanden, K. W., & Ofstad, R. (2019). Near-infrared spectroscopy detects woody breast syndrome in chicken fillets by the markers protein content and degree of water binding. Poultry Science, 98(1), 480–490. http://dx.doi.org/10.3382/ps/pey351.
Yoon, S. C., Bowker, B. C., Zhuang, H., Bowker, B. C., & Zhuang, H. (2016). Toward a fusion of optical coherence tomography and hyperspectral imaging for poultry meat quality assessment. Electronic Imaging, 28, 1–5. https://doi.org/10.2352/ISSN.2470-1173.2016.14.IPMVA-380.
Zhou, Y., Wang, F., Zhou, W., & Wu, Y. (2021). Feasibility study of detecting plum’s early mechanical injury based on optical coherence tomography and cell morphological parameters. Journal of Food Processing and Preservation, 45(9), e15664. https://doi.org/10.1111/jfpp.15664.
Zhou, Y., Wu, D., Hui, G., Mao, J., Liu, T., Zhou, W., Zhao, Y., Chen, Z., & Chen, F. (2018). Loquat bruise detection using optical coherence tomography based on microstructural parameters. Food Analytical Methods, 11(10), 2692–2698. https://doi.org/10.1007/s12161-018-1246-6.
Zhou, Y., Wu, Y., & Chen, Z. (2022). Early detection of mold-contaminated maize kernels based on optical coherence tomography. Food Analytical Methods, 15(6), 1619–1625. https://doi.org/10.1007/s12161-021-02205-y.

No competing interests reported.

Download PDF

Editorial decision: Major revision
15 Oct, 2023
Reviews received at journal
14 Aug, 2023
Reviewers agreed at journal
05 Aug, 2023
Reviewers invited by journal
05 Aug, 2023
Editor assigned by journal
04 Aug, 2023
Submission checks completed at journal
03 Aug, 2023
First submitted to journal
02 Aug, 2023

You are reading this latest preprint version

Nondestructive Assessment of Woody Breast Myopathy in Chicken Fillets Using Optical Coherence Tomography Imaging with Machine Learning

Status:

Version 1

Abstract

Figures

Introduction

Materials and Methods

Chicken Fillets

OCT Device and Data Acquisition

OCT Data Processing and Analysis

Feature Extraction

Machine Learning Classification Models

Results and Discussion

Visualization and Observations

Histogram Analysis

Woody Breast Condition Classification

Conclusions

Declarations

References

Additional Declarations

Status:

Version 1