Rat preparation and intrathecal drug delivery.

While preparing animals for the implantation of the PE–10 intrathecal (i.t.) catheter (Becton Dickinson, Sparks, MD, USA), male Wistar rats (320–400 g) were anesthetized with phenobarbital (65 mg/kg, intraperitoneally; Sigma) and implanted with one i.t. catheter (8.5±0.5 cm) via the atlantooccipital membrane down to the lumbar enlargement (L1-L2) of the spinal bony structure. We used a polyethylene tube (8.5±0.5 cm in length, 0.008 in. ID, 0.014 in. OD) and a 3.5 cm silastic tube to construct intrathecal catheters. The polyethylene tube was inserted into the silastic tube and sealed with epoxy resin and silicon rubber. The end of the catheter was then externalized and fixed to the dorsal aspect of the head. The i.t. catheter was linked to a mini-osmotic pump and used to infuse morphine (15 μg/h), saline (1 μl/h), gabapentin (15 μg/h, Sigma), or morphine (15 μg/h) plus gabapentin (15 μg/h, Sigma) for five days. The morphine and saline groups were used for training the prediction model, and the gabapentin group is used for independent test of the model. After catheterization, all rats were returned to their home cages for rest. Each rat was housed individually and maintained on a 12 h light/dark cycle with food and water available. Rats with neurological deficits were excluded.

Anti-nociceptive test.

The anti-nociceptive test is adjusted from previous studies [3,7]. One group of 15 rats received a constant morphine infusion for five days for tolerance development, while a control group of 15 rats received a constant saline infusion. An additional group of 10 rats received a co-infusion of morphine and gabapentin to evaluate the gabapentin effect on morphine tolerance development. All drug infusions were administered at a rate of 1μl/h through a mini-osmotic pump (Alzet, Cupertino, CA) implanted in the interscapular region. The development of morphine tolerance was assessed by a tail-flick latency test. Tail-flick latency using a hot water immersion test (52±0.5°C) was documented prior to drug infusion and once daily for 5 days after infusion began. The baseline tail-flick latency was 2±0.5 s and 10 seconds was allocated as the cutoff time to prevent tail injury. Rats were placed in plastic restraints during drug injection and the anti-nociception assay. The morphine analgesic effects were converted to a percentage of the maximal possible effect (%MPE) using the formula: MPE = [(test latency—baseline latency)/(maximum latency—baseline latency)]×100. We monitored the effect of gabapentin on morphine tolerance using the tail-flick latency test, and also used the immunohistochemistry method to detect the level of neuroinflammation related to the morphine and gabapentin co-infusion. An anti-nociception response curve was constructed for each study group.

The tail-flick latency test indicated the group receiving chronic intrathecal infusion of morphine developed morphine tolerance. The behavioral tail-flick assay on the 10 gabapentin-treated rats shows a significant attenuation of morphine tolerance (see the Results and Discussion section).

Spinal cord immunostaining sections and image acquisition.

Following the completion of the experiments, all the rats were sacrificed by exsanguination under isoflurane anesthesia (ABBOTT, Abbott Laboratories Ltd., Queenborough, Kent, England). A laminectomy was performed at the lower level of the twelfth thoracic vertebra, and the L5–S3 segment of the spinal cord was removed and embedded in an optimal cutting temperature compound (Sakura Finetec Inc., USA) for immunohistochemistry.

Spinal cord sections (5μm) were fixed by soaking in ice-cold acetone/methanol (1:1) for 5 min. After three washes in ice-cold phosphate-buffered saline (PBS), sections were categorized by incubation overnight at 4°C with the FITC-labeled mouse monoclonal anti-glial fibrillary acidic protein (GFAP) antibody for astrocytes (Molecular Probes, Oregon, USA), the FITC-labeled mouse monoclonal anti-rat CD11b/c antibody for microglia (OX42; Serotec, Oxford, UK), and the FITC-labeled mouse monoclonal anti-neuronal nuclei for neurons (Chemicon, Temecula, CA) diluted in 1% normal goat serum in PBS. After three PBS rinses, 400x magnification images were taken using an Olympus BX 50 fluorescence microscope (Olympus Optical, Tokyo, Japan) and a Delta Vision disconsolation microscopic system operated using SPOT software (Diagnostic Instruments Inc., USA). For each rat, three images were produced for the three cells astrocytes, microglia, and neurons.

Image preprocessing.

The original immunostaining images are recorded in RGB format. The following preprocessing procedure was applied to the original images of cells in the spinal cord: 1) the color images are converted into gray scale and binary formats; 2) the scale of intensity is normalized by linear transformation into [0, 65,535]; 3) the image size is down-sampled from 1600×1200 to 800×600; 4) all objects smaller than 50 pixels are removed, and 5) background noise is removed by using the Otsu’s method [22]. Fig 2 shows example images of the saline controls and morphine-tolerant rats with neuroinflammation of cells in the spinal cord and their corresponding preprocessed images.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Typical images of control and morphine-tolerant cells in the spinal cord and their preprocessed images.

(A) Astrocytes of the saline control. (B) Astrocytes with morphine tolerance. (C) Microglial of the saline control. (D) Microglial with morphine tolerance. (E) Neurons of the saline control. (F) Neurons with morphine tolerance. Scale bar represents 25μm.

https://doi.org/10.1371/journal.pone.0139806.g002

Image Feature extraction.

Image-based high-content screening (HCS) has been shown to be a promising phenotypic screening approach for drug discovery. However, the majority of previous HCS-based studies (60–80%) made use of only one or two image-based features measured from each sample and disregarded the distribution of those features among each cell population [20]. Feature extraction and feature selection methods from HCS images play important roles in designing computer-aided image diagnosis systems [21]. To our best knowledge, no previous studies have reported effective feature identification methods for identifying the phenotypic change of neuroinflammation in the HCS images of cells in the spinal cord.

We comprehensively investigate numerous potential features that can be extracted from the readout of the HCS images to characterize phenotypic changes of cells with neuroinflammation activity. All the features can be categorized into five classes according to their properties (i.e., moment, texture, intensity, morphology, and frequency) which describe local or/and global properties of multiple cells in a highly multiplex immunostaining image. The design of the HCS-Morph method accounts for both the properties of individual cells and also the distribution properties of multiple cells in the whole image.

This work uses 21 feature sets belonging to the five classes as follows: 1) moment: geometric moments [23], Legendre moments [24], Tchebichef moments [25], Krawtchouk moments [26], Zernike moments [27], pseudo-Zernike moments [24], radial Tchebichef-Fourier moments [27], and Fourier-Mellin moments [28]; 2) texture: gray scaled co-occurrence matrix (GLCM) [29] and Haralick texture [29,30]; 3) intensity: histogram and relative histogram; 4) morphology: square model shape matrix [31], polar model shape matrix [32], and region properties [33], and 5) frequency: subcellular location feature SLF1 [34], generic Fourier descriptors [35], Granularity spectrum [36], Gabor filter [34], Wavelet energy (Daubechies4, Haar) [34], and Fourier transform [37]. The total number of candidate features in the 21 feature sets is N = 5,033, which can be roughly categorized into three classes as follows.

1. Neuronal features: Five features are commonly used in the literature, including Mean Cell Intensity, Cell Number, Total Cell Area, Largest Cell Area, and Largest Cell Diameter. Some neuronal morphology quantification methods can be used to extract the neuronal features [38,39]. These features are extractable, interpretable, and understandable in observing morphological changes of neuron cells [10,40,41].
2. Interpretable features: Interpretable features reveal their meanings directly from the whole images of cells such as intensity variation, morphology change, and texture homogeneity. They are directly computable and human-interpretable, e.g., the number of pixels in a specific range of intensity (“Pixel Area”) and the occurrence (“Texture Area”) and variation (“Texture Area Variation”) of pixel pairs that satisfy a specific condition.
3. Computational features: These features describe global image properties with complicated mathematic functions such as Polar-coordinate-based moments, Cartesian-coordinate-based moments, Haralick texture, Fourier descriptors, and wavelet energy. The features are not easily observed directly but may effectively respond to phenotypic changes of cells in the spinal cord.

Image feature selection.

In developing an effective method for identifying HCS-based features to predict neuroinflammation in morphine tolerance, this work considers the following facts. First, because no prior knowledge is utilized, many potential features (N = 5,033) in the HCS images are extracted for comprehensive investigation. Second, simultaneous selection of multiple features generally provides better prediction accuracy than the combination of features individually selected by univariate analysis [42]. Third, biologists and clinical researchers require not only high prediction accuracy but also high feature interpretability. This work proposes a coarse-to-fine feature selection approach to obtain a small set of informative and easily-interpretable features for predicting neuroinflammation in morphine tolerance. The proposed neuroinflammation feature selection method consists of the two steps: univariate feature selection and multivariate feature selection, as described below. The same feature selection method is applied to each type of the three cells in the spinal cord.

1. Univariate feature selection. The p-value (P) for each of N candidate features is calculated using t-test analysis to identify the significant features for distinguishing neuroinflammatory and non-neuroinflammatory cells. The feature set of size n obtained from N features in this coarse step consists of three classes: 1) the best n₁ features according to p-value, 2) the best n₂ features considering both p-value and human interpretability, and 3) n₃ neuronal features reported in the literature to characterize the cells in the spinal cord. In this work, n₁ = 15, n₂ = 20, n₃ = 5, and n = 40. The n₁ features may come from the sets of neuronal, interpretable, and computational features. The n₂ interpretable features are selected with human intervention. The above-mentioned five neuronal features of morphology (n₃ = 5) are adopted.
2. Multivariate feature selection. Multivariate feature selection identifies a small set of m features responding to the phenotypic change by maximizing prediction accuracy for neuroinflammation. An intelligent bi-objective combinatorial genetic algorithm (IBCGA) is used to solve the combinatorial optimization problem of C(n, m) having a huge search space of size C(n, m) = n!/(m!(n-m)!). The IBCGA uses an intelligent genetic algorithm (IGA) [43] with an inheritance mechanism [42] to efficiently search for the solution S_r+1 to C(n, r+1) by inheriting the good solution S_r to C(n, r). The IGA algorithm based on orthogonal experimental design uses a divide-and-conquer strategy and a systematic reasoning method instead of the conventional generate-and-test method to efficiently solve the large-scale combinatorial optimization problem. The IBCGA algorithm solves the combinatorial optimization problem by minimizing the number of selected features and maximizing the fitness function f(S) defined as the prediction accuracy of 10-fold cross-validation (10-CV) for the candidate solution S. The design of the prediction model is to simultaneously determine the optimal parameter settings of the support vector machine (SVM) and feature selection. The input for the SVM-based model design procedure is a training dataset of two-class images (neuroinflammation and non-neuroinflammation) and the output contains a set of m selected image features and an SVM classifier with associated parameter settings of γ and C. A radial basis kernel function exp(-γ||x_i—x_j||²) is adopted, where x_i and x_j are training samples, and γ is a kernel parameter [44]. In this work, γ ∈{2⁻⁷, 2⁻⁶, …, 2⁸} and C∈{2⁻⁷, 2⁻⁶, …, 2⁸}. The input and output of the used IBCGA algorithm are the training samples of n features and m selected features, respectively.

The IBCGA with the fitness function f(S) can simultaneously obtain a set of solutions, S_r, where r = r_start, r_start+1, …, r_end in a single run. In this work, the parameter settings are r_start = 2, r_end = 30, N_pop = 60, p_c = 0.8, p_m = 0.05, and Gmax = 60. The customized IBCGA algorithm for the image feature selection is given below.

1. Step 1:. (Initiation) Randomly generate an initial population of N_pop individuals. All the n binary genes f_i have r 1’s and n-r 0’s where r = r_start.
2. Step 2:. (Evaluation) Evaluate the fitness values of all individuals using f(S).
3. Step 3:. (Selection) Use a conventional tournament selection that selects the winner from two randomly selected individuals to form a mating pool.
4. Step 4:. (Crossover) Select p_c·N_pop parents from the mating pool to perform orthogonal array crossover [43] on the selected pairs of parents where p_c is the crossover probability.
5. Step 5:. (Mutation) Apply a conventional mutation operator to the randomly selected p_m·N_pop individuals (except the best individual) in the new population where p_m is the mutation probability.
6. Step 6:. (Termination test) If the stopping condition of performing Gmax generations for obtaining the solutions S_r is satisfied, output the best individual as S_r. Otherwise, go to Step 2.
7. Step 7:. (Inheritance) If r < r_end, randomly change one bit in the binary genes f_i for each individual from 0 to 1; increase the number r by one, and go to Step 2.
8. Step 8:. (robustness) Perform Steps 1–7 for R (= 30 in this work) independent runs and obtain the best one of R solutions.

The best solution can be determined by considering the most accurate one S_a with the highest fitness value or the robust one S_r with the highest appearance score [45]. The appearance score considers both the fitness value and the mean number of times for individual features selected in the R runs. In this work, the solution S_r was adopted.

Prediction platform for neuroinflammation.

Neuroinflammation activity plays an important role in the development of morphine tolerance [5,46]. The role of neuroinflammation in the development of morphine tolerance and the effect of the TNF-α inhibitor, etanercept, in the attenuation of morphine tolerance are explained in the review article [2]. The proposed HCS-Morph method uses an SVM-based classifier with informative image features to predict images of astrocytes, microglia, and neurons in the spinal cord. Informative image features should have good prediction ability to respond to phenotypic changes caused by neuroinflammation. Interpretability of the informative features should be also considered to gain an insight into phenotypic changes of cells in the spinal cord. Moreover, interpretable features relating to neuroinflammation can moderately cope with the problem of high-level background noise. A related study of the identification of phenotypic changes in multi-neuron images upon drug treatments with high-content screening can be found in our previous work [21]. The present work aims to identify sets of HCS-based neuronal features to establish an accurate prediction platform for neuroinflammation by solving the bi-objective optimization problem. To make the best use of all available samples, 15 controls and 15 morphine-tolerant rats were used to train the HCS-Morph model for establishing the neuroinflammation prediction platform.

Since neuroinflammation prediction has relatively lower cost (in terms of rats, experimental material, time, etc.) than traditional experimental approaches, the platform can benefit drug screening processes through minimizing the need for drug-treated rats. Once potential drugs are discovered, conventional experiments can be used for further validation of the drugs. Therefore, the computer-aided image diagnosis system based on HCS-Morph with accurate prediction of neuroinflammation can greatly facilitate the development of tolerance therapy with anti-neuroinflammatory drugs.