• Nenhum resultado encontrado

3.2 Wound image classification

3.2.4 Similar studies

Finally, studies with concern to post-operative wounds are described. Even though their employed datasets do not focus specifically on cardiothoracic surgical site’s, the wound type is the most similar with this work.

Zenget al.[108] proposed a post-operative wound analysis by automatically detecting the wound followed by the recognition of an infectious status. Wound segmentation is obtained by combining edge detection, color normalization and skin color detection (separation of skin area and non-skin area). A high accuracy of 92.96% was achieved for this step of wound detection. However, the sensitivity was 75.82% and the specificity was 97.61%. They separated the image classes into three infectious statuses of wounds:

C H A P T E R 3 . L I T E R AT U R E R E V I E W

swelling, blood region and necrosis. Histogram based features and GLCM features were extracted and trained on a SVM model. Furthermore, the accuracies of all statuses were higher than 76%, averagely. This paper had a big disadvantage, the dataset was composed of only six images, which makes the traditional segmentation algorithm have a higher accuracy. These parameters, like thresholds can diverge if a bigger dataset was used.

Hsu et al. [109] presented an automatic way to perform wound segmentation and infection assessment after surgical operations. The wound segmentation consisted of an edge-based self-adaptive threshold method to remove the non-wounded areas from the original images. For edge detection, the authors applied Canny edge detection and connecting component labeling to identify each object present in the image. Overall, it achieved an accuracy of 89.04% with a TP rate of 76.44%. For assessing the wound infec-tion, scale-invariant feature transform (SIFT) and cross-shaped features were considered.

However, after feature selection, the pixels belonging to the cross-shaped regions were grouped into multiple clusters by a threshold algorithm. These formed clusters were then merged into several ROIs, to extract a feature vector based on the values of hue, pixel histogram and mean value of histogram. For anomaly detection, it showed an 87.31%

accuracy value and for symptom assessment, 83.58% of accuracy value. In this research, 134 surgical wound sample images, including chest, abdomen, back, hand, and podiatry wounds, were processed by robust image segmentation and SVM-based wound infection assessment.

In [110], the objective was to classify wound infection features using DL. Wuet al.

developed a drawing tool to mark the wound regions and wound features. The wound ROI detection was obtained with an alteration of a ResNet50 architecture pre-trained on Pascal VOC dataset. The obtained IoU was 77.5%. To classify wound infection a one deep CNN (Xception) was used to perform a binary classification and three classifiers, SVM, RF and gradient boosting were trained, to compare the both methods. Their evaluation achieved an accuracy of 79.5%, recall of 77.1%, precision of 82.7%, F1 of 79.4% and AUC score of 83.3% for the DL method, while the classifiers achieved lower AUC scores (SVM - 44.4%, RF - 67.1%, gradient boosting classifier - 66.9%).

In conclusion, ML algorithms are widely used for wound classification due to its sim-plicity and good performance. However, DL and hybrid approaches started to surface when the extraction of hand-crafted features for certain problems gave a bad perfor-mance.

38

4

Methodology

In this chapter the proposed methodology is described for the understanding of how the wound image analysis system was developed and implemented in the established dataset. The proposed pipeline is presented and then divided into two main subsections:

wound segmentation and wound classification.

4.1 Dataset

The dataset is composed by surgical site RGB images from 34 cardiothoracic surgery patients of Hospital de Santa Marta. The images in the database belong to patients with ages between 24 and 83, with the highest percentage belonging to the age ranging from 60-70. Additionally, the gender distribution was pretty evenly, which makes the dataset more realistic and relevant. Several types of cardiothoracic procedures were performed, 26%

of CABG, 50% of valve procedures, 10% of a combination of both and other remaining types of surgery.

The photographs correspond to the evolution of each patient’s surgical wounds during a 30-day follow-up. Everyday, patients or family members take an image of the surgical sites and send it through the telemonitoring platform for prior evaluation by the clinical team. Initially, there were a total of 1443 images collected by the front or back camera of a Xiaomi Mi A2 Lite smartphone. As so, the images’ resolution varied among samples, but the majority had a 1920x1080 pixel resolution. The acquisition protocol was not defined, resulting in images with several different conditions, such as differences in illumination, patient position, orientation and background. For this reason, images with very poor illu-mination conditions or heavily blurred together with unrelated images taken by accident, were eliminated from the dataset. After the image removal process, the final dataset had a total of 1337 images.

The acquired images could have three types of wounds: chest wound (WC), drainage wound (WD) or a leg wound (WL). An example of each wound type is illustrated in Figure 4.1. Another suture incision near the clavicle also appeared in the images, however this type of wound was so scarce that it was not considered in the dataset. Each image could

C H A P T E R 4 . M E T H O D O L O G Y

have more than a wound type and more than one wound of the same type. The total amount of images had 77.49% of WC, 67.02% of WD and 20.79% of WL.

(a) Chest wound (b) Drainage wounds (c) Leg wound

Figure 4.1: Example of the wound types present in the dataset.

In terms of the classification problem, the wounds can be categorized into binary labels, where 0 indicates the wound does not have any concerning alterations and 1 means that an alteration is present in the wound. An example of two wounds belonging to the same patient, with and without alterations is illustrated in Figure 4.2. Only 10.70%

of images had displayed wounds with alterations, as so with such scarce in the positive class, it resulted in an imbalance dataset. Even with a considered large dataset, since each patient took in average 30 photographs, the images correspond to the same wounds but with few minor alterations. Therefore, the dataset is not well rounded and diverse. The reduction of the dataset dimensions throughout the pipeline, is exhibited in Figure 4.3.

(a) Altered wound (b) Non-altered wound

Figure 4.2: Example of the evolution of a patient’s WC. The altered wound (a) exhibits a clear disunion of the suture’s borders that decreased with time resulting in (b) that shows a minor distance between the borders.

40

4 . 1 . DATA S E T

Figure 4.3: Reduction of the dataset’s dimensions. The full dataset, with a total of 1443 images, was reduced by a cleaning step of excluding poor quality images. After the pre and post-processing steps of both segmentation and classification models, 3146 wounds remained, which were distributed into the three established wound types.