**1. Introduction**

Laryngeal cancer is the second most frequent malignant tumor of the head and neck region [1]. The vast majority of primary laryngeal cancers are Squamous Cell Carcinomas (SCC) arising from the epithelial lining of the larynx, mostly as a result of tobacco and alcohol consumption. A total of 40% of these cancers are diagnosed at an advanced stage, which is associated with a poorer prognosis and quality of life [2]. The early diagnosis of laryngeal cancer is crucial to reduce patient mortality and preserve vocal fold function.

Specific changes in the morphology and three-dimensional orientation of the vocal fold's sub-epithelial blood vessels have proved to be associated with the development of benign and malignant laryngeal lesions. Several approaches have been proposed to describe and classify these vascular changes. Among the complex classification systems proposed by [3] and [4], the European

Laryngological Society (ELS) introduced a simplified classification that divides vascular changes into longitudinal and perpendicular classes [5,6]. Longitudinal Vascular Changes (LVC) spread along the length and width of the vocal fold and can be observed in all kinds of benign or malignant lesions. On the contrary, Perpendicular Vascular Changes (PVC) develop perpendicularly towards the mucosa, as a result of neoangiogenesis in laryngeal Papillomatosis, pre-malignant and malignant histopathologies.

The endoscopic detection and evaluation of vascular changes can provide complementary diagnostic information for clinicians to detect and differentiate between benign and malignant laryngeal lesions [7]. As a minimally-invasive endoscopic technique, Contact Endoscopy (CE) can provide real-time visualization of cellular and vascular structures of the laryngeal mucosa [8,9]. For the purpose of detecting and evaluating superficial vascular changes, several enhanced endoscopic techniques such as Narrow Band Imaging (NBI) have been combined with CE to ease the detection of vascular changes [10]. The use of enhanced CE showed promising results in the assessment of vascular patterns followed by indicative of various laryngeal pathologies [4,11,12].

Clinicians can receive useful information about the type and suspected histopathology of laryngeal lesions by evaluating LVC and PVC in enhanced CE images; however, it is a challenging task for them. There are similarities between vascular patterns of benign and malignant laryngeal lesions. The PVC with wide-angled turning points, as observed in laryngeal Papillomatosis can be difficult to distinguish from PVC with narrow-angled turning points, as observed in pre-malignant and malignant histopathologies [5,12–14]. Hence, the interpretation of vascular patterns in enhanced CE images requires an extensive learning curve from the clinicians to reduce the risk of subjective evaluation that can cause potential problems in differentiation between benign and malignant laryngeal lesions [4,10,12,15,16].

In this study, we first aimed to present the results of manual versus automatic classification of benign and malignant laryngeal lesions based on the vascular patterns in CE-NBI images. We then evaluated the issues of manual classification and subsequently showed how a computer-based approach can assist the clinicians to overcome these problems. A manual and an automatic classification approach were defined to conduct this evaluation. In the manual approach, six experienced and less experienced otolaryngologists individually evaluated PVC and LVC in CE-NBI images of patients and classified them into benign and malignant groups. An updated version of the algorithm proposed in [17,18] with 24 features and four supervised classifiers has been used to classify CE-NBI images into benign and malignant groups. The results of the two approaches were compared in terms of classification sensitivity and specificity. The potential of an automatic approach to assist the clinicians is presented through two evaluation strategies.

#### **2. Material and Methods**

### *2.1. Data Acquisition*

CE-NBI images were extracted from video scenes of adult patients who received a microlaryngoscopy for benign, pre-malignant or malignant lesions of the vocal folds. Video scenes were captured using an Evis Exera III Video System with integrated NBI-filter (Olympus Medical Systems, Hamburg, Germany) and a rigid 30-degree contact endoscope (Karl Storz, Tuttlingen, Germany) with a fixed magnification of 60 to have a fixed camera–tissue distance. For each video scene, we selected the time intervals where the video quality was good enough to visualize the vessels. Then, one in every ten frames was extracted from the selected intervals in JPEG format images (1008 × 1280 pixels) to have unique and non-redundant CE-NBI images.

#### *2.2. Dataset Generation*

The CE-NBI dataset included 1632 extracted images of 68 patients. The patients' data were pseudonymized. Based on the WHO classification [19], histological diagnoses were used to label images as belonging to a benign or a malignant class. Table 1 shows the histopathologies with the number of patients and images used for the generation of the dataset.

Two image subsets were created from the CE-NBI dataset. The *Subset I* included a series of two to five randomly selected CE-NBI images of each patient—total of 336 images, ≈ 20% of the dataset. The *Subset II* included the rest of the CE-NBI images—a total of 1296 images, ≈ 80% of the dataset, and was used as the training set of the automatic approach. The *Subset I* was evaluated by the otolaryngologists in the manual approach and then used as the testing set for the automatic approach. Figure 1 presents some examples of CE-NBI images with LVC and PVC belonging to the generated dataset.


**Table 1.** Histopathologies used for the generation of the dataset.

**Figure 1.** Examples of Longitudinal Vascular Changes (LVC) and Perpendicular Vascular Changes (PVC) in Contact Endoscopy (CE)-Narrow Band Imaging (NBI) images with different histopathologies: (**a**) Reinke's edema, LVC; (**b**) polyp, LVC; (**c**) amyloidosis, LVC; (**d**) severe dysplasia, PVC, (**e**) carcinoma in situ, PVC; (**f**) Squamous Cell Carcinomas (SCC), PVC.
