This retrospective study has been approved by the Institutional Review Board (IRB) of the authors’ affiliated institutions. Proper informed consent was obtained from all patients. The patients included in this study were determined with the following inclusion criteria: patients over 18 years old; with confirmed ACL acute and chronic injury; submitted to ACL reconstruction surgery. The following exclusion criteria were adopted: patients under 18 years old; with any autoimmune disease that causes joint impairment; with positive sorology for: Hepatitis B or C, HIV, HTLVI / II or Chagas disease. After this stage, patients were conducted to the plasma autologous collection necessary to produce the platelet-derived hormones.
In our study, we utilized different scores to demonstrate knee conditions before and after (3 months) the surgery for both our groups (19-23). Those scores are as follows: knee circumference, pain, flexion, IKDC index (International Knee Documentation Committee) (24), and Lysholm Score (25).
The patients passed through clinical and laboratory triage to test blood transmissible diseases to determine their eligibility according to the autologous donation legislation No. 158 of February 4, 2016 (26). Approved candidates for clinical screening were referred for autologous whole blood donation, with subsequent segregation of blood components: red blood cells (RBC), Frozen Fresh Plasma (FFC) and Platelet Concentrate (PC) (27, 28).
The statistical analysis was performed using PRISM 7 software (GraphPad software, Inc., 2016). Significant differences in the mean values of scores described above were determined using the paired and not paired Student T-test. We considered them statistically different when the p-value was < 0.05.
2.1 PRP Production and Storage
Following the donation, the blood components were processed according to institution standards. The RBC unit was stored in an appropriate cooler. The FFP was frozen in an automated temperature decay system in Thermogenesis® equipment and then transferred to a freezer at -80 ° C. The PC was in continuous agitation until the preparation of the Platelet Derived Hormones (PDH) (27).
The FFC and PC units of each patient were sent to the Cell Engineering laboratory of our institution for the production of PDH. After processing, the cryoprecipitate and the PDH were ready for application(27).
2.2 Study Group
After the selection, 34 patients were included in this study. They were divided into 16 patients as the control group (mean age: 32 ± 7) and 18 patients (mean age: 29 ± 8) as the experiment group. The control group was composed of patients submitted to the operative procedure of ligament reconstruction with an arthroscopic technique using the gracilis and semitendinosus tendons. It also used endobutton in the femoral tunnel and metal interference screw in the tibia. In the control group, there was no application of PRP (29).
The experimental group was composed of patients with surgical indication who underwent the same procedure performed for the control group. The difference was that prior to the closure of the operative field, the patient received the application of 4 ml of PRP in loco (28). All the patients included in this study performed magnetic resonance imaging of the knee three months after surgery.
MRI was performed using a 3T scanner (Magnetom Verio, Siemens, Erlangen, Germany) with a 8-channel-knee coil. Turbo-spin-echo (TSE) DP FAT SAT sequences (FOV: 170 × 170 mm; matrix: 384 × 384; slice thickness 2.5 mm; Flip angle: 150º; GAP 1; TR/TE = 3550/44; NEX: 2) were acquired on the sagittal view. Total sequence acquisition time was approximately 4 minutes. In this study we utilized MRI T2-weighted images for the subsequent steps described below.
2.3 Texture Analyses (TA)
2.3.1 TA Features Extraction
Textures are attributes present on images that correspond to a visual pattern or arrangement of structure, usually related to the distribution of pixels. Texture usually contains very significant information about image content which makes it widely used in image processing (30). Texture analysis (TA) refers to the characterization of regions based on their texture content. TA applied in medical imaging provides a tool to classify images, or to differentiate between healthy and pathological tissues (31).
The TA features extraction was performed in Matlab® software R2017a. First, we selected the RoIs in sagittal slices of MRI T2-weighted images. The slice with the largest ACL longitudinal section visible in the image was selected (figure 1 – A). The RoIs delimited the region of the ACL of each patient (figure 1 - B). RoI positioning was performed by two experienced radiologists. From the delimited RoIs, 62 texture parameters were extracted, including Gray-Level Co-occurrence Matrix (GLCM) and Gray-Level Run-Length (GLRL) (32, 33). GLCM contains the second-order statistical information of neighborhood pixels of an image. GLRL provides a dimensional matrix that calculates the number of adjacent pixels that have the same gray intensity in a particular direction (32, 33).
2.3.2 Machine Learning Analyses and Algorithms
After the previous we transferred the dataset to the Orange Canvas (v 3.18) software. Orange contains a powerful library that includes a selection of machine learning methods, processing methods, and sampling techniques. We used Orange Canvas to perform the analyses in all our extracted texture features dataset (34). First, the dataset was divided into two independent parts: a training set (75% of the total sample) and a test set (the remaining 25%), with a 10 fold cross-validation. Different machine learning classifiers were utilized with our training set, to determine the best combination of features to achieve the highest accuracy. We used the following learning classifiers:
Logistic Regression (LR) which is used to analyze the relationship between continuous or categorical predictive variables (explanatory or independent), and a categorical outcome that produces binary variables responses (dependent). LR estimates the probability of the dependent variable to assume a certain value as a function of known variables (35).
Stochastic Gradient Descent (SGD) which is a standard algorithm to optimize complex functions iteratively. SGD has a high impact on machine learning with its optimization method for unconstrained problems. It approximates the true gradient by considering a single training example. The algorithm works iteratively over the training examples updating the model parameters with each iteration (36).
Naive Bayes (NB) which is based on the Bayes’ theorem and the maximum posterior hypothesis, assuming that the effect of an attribute on a given class is independent of the values or other attributes called “conditional independence”. The classification does not require an accurate probability estimate as long as the maximum probability is assigned to the correct class (18).
Gain ratio and Gini index were used to rank all features according to their correlation with each class (Raileanu and Stoffel, 2004, Stoffel and Raileanu, 2001). Thus, from the 62 texture features we selected the five features that achieved the highest scores for classification within each machine learning classifier. To determine how efficiently the models classified our groups, we utilized quantitative parameters such as the area under the Receiver Operating Characteristic curve (ROC), accuracy (CA), F-score (F1), Precision and Sensitivity.