Value of high-speed videoendoscopy as an auxiliary tool in differentiation of benign and malignant unilateral vocal lesions

The study included 175 patients hospitalized in the Department of Otolaryngology, Head and Neck Oncology of the Medical University of Lodz from 09.2020 to 12.2022. The control group consisted of 50 patients: 33 women and 17 men. The inclusion criteria for this group were: > 18 years of age, no dysphonia symptoms at the time of the examination, no history of dysphonia, and no structural or functional abnormalities found in endoscopic examination. Subjects in this group were between 19 and 83 years old, with an average age of 41.52 ± 17.44 years.

The study group consisted of 125 patients. Inclusion criteria for the study group were: > 18 years of age; current dysphonia, presence of unilateral organic lesion in the glottis. Subjects with dysphonia were later divided into two groups. The benign lesions group included 85 participants diagnosed with benign vocal lesions (including polyps, cysts, nodules, and unilateral Reinke’s edema). This group consisted of 58 women and 27 men, aged from 22 to 87 years, with the mean age of 53.28 ± 13.3 years. The second group included 40 patients (30 men and 10 women) diagnosed with early glottic cancer. Participants in this group were aged from 42 to 85 years, with an average age of 67.45 ± 7.01 years. All patients with organic lesions underwent transoral laser microsurgery (TOLMS), resulting in histopathologic confirmation of the initial diagnosis. All the malignant lesions were squamous cell carcinoma G1 or G2, classified clinically as T1 involving one vocal fold.

The exclusion criteria for both groups were inability to obtain recordings with sufficient visualization quality for analysis, and organic lesions involving both sides of the larynx.

Approval for this study was granted by the Ethical Committee of the Medical University of Lodz (no. RNN/96/20/KE 08/04/2020), and all patients gave written informed consent to participate in the research.

First, the patients were assessed by an ENT physician who made the initial diagnosis. After that, every patient was subjected to examination by means of a rigid endoscope paired with HSV camera and laser illumination. The examination equipment was supported by a computer with dedicated software allowing for the recording of video signal and storing it for further analysis. Initially, the camera was set at a recording speed of 24 frames per second (fps) to center the video sensor on the glottis and perform a primary assessment of the larynx. Next, the camera was set into high-speed recording mode, and multiple HSV sequences were captured. For this study recording speed was 3200 fps. Each sequence contained 2000 fps, with the full-time length of a single sequence of 625 ms (Malinowski et al. 2021). More details on the technical aspects of HSV recording can be found in our previous publication. Later, quality of the recordings was assessed, and the best and most representative films were selected for each patient for further assessment.

Detailed kymographic analysis was performed to quantitatively assess vocal fold vibrations. As described by Yamauchi et al. sets of data from HSV kymography can be classified as three dimensions of HSV—mediolateral (x), longitudinal (y), and temporal (t), as shown in Fig. 1 (Yamauchi et al. 2021). The first two—the x- and y-axis—describe spatial properties of glottal oscillations and can be assessed subjectively (visual-perceptual rating), by laryngotopography and by means of HSV (Sakakibara et al. 2010; Tsuji et al. 2014; Yamauchi et al. 2015). The temporal aspect can be evaluated either on its own (by analysis of glottal area and width waveform) or along with mediolateral dimension—using digital kymography (Schlegel et al. 2020; Kist et al. 2021a; Yousef et al. 2023). In this study, to compare involved and healthy vocal folds, we focused on the spatial aspect of HSV analysis (x- and y-axis), resulting in the calculation of STV parameters including amplitude measures, glottal dynamic characteristics, and symmetry measures. Those parameters describe the movement of vocal folds along and perpendicularly to the glottal axis.

First, the software automatically selected a part of the recording for further analysis. Contrary to LVS, in the case of HSV recordings, manual stabilization and centering of the image for each frame was not required, because, in such a short recording time (625 ms), the movement of the endoscope tip is negligible. Later the vocal folds edges were identified semi-automatically—the examiner marked them in five points on a single representative frame, and based on those points, the software indicated them on all frames and along the whole length of the glottis. The examiner was then able to verify and adjust the automated detection in case of any discrepancy. In the final step, the software calculated a set of STV parameters describing characteristics of the movement of vocal folds during the glottal cycle, specifically: amplitude measures, glottal dynamic characteristics, and symmetry measures. Amplitude measures include four parameters: average amplitude (AmpAvg), the average amplitude of the middle third of the glottis (AmpAvg_2/3), the average amplitude of the involved vocal fold (AmpInvolvedAvg), and the average amplitude of the healthy vocal fold (AmpHealthyAvg)—their values indicate the average resultant amplitude of vocal fold movement for the whole glottal gap, its middle third part, involved, and healthy vocal fold, respectively. This parameter is calculated as an averaged value for all points along the length of respective vocal fold. Whenever a parameter below has a suffix “_2/3”, it relates to the middle third part of the glottal length. We decided to include those parameters as the middle part of the glottis which is known to be the location of the maximum amplitude of oscillations, and disruptions in this area tend to have a larger impact on phonation. In our study, organic lesions were always unilateral, so based on the location of the lesion, we compared both folds in a series of analyses. For this purpose, we used AmpInvolvedAvg and AmpHealthyAvg to compare vibrations of healthy and involved vocal folds. An exception was made in the statistical analysis with regard to these parameters—comparison was performed only in benign and malignant lesion groups. Among glottal dynamic characteristics, four parameters are calculated: average open quotient (OQAvg), OQAvg_2/3, relative glottal gap area (RGGA), and non-opening. OQAvg and OQAvg_2/3 describe the average open quotient for the whole glottal gap and its middle third part, respectively, indicating the ratio of the glottal opening phase to the whole length of the vocal cycle. RGGA defines the ratio of minimal to maximal area of the glottis during the cycle, while non-opening indicates part of the glottis without opening during the cycle. Symmetry measures include five parameters: average amplitude asymmetry (AmplAsymAvg), AmplAsymAvg_2/3, average phase asymmetry (PhaseAsymAvg), PhaseAsymAvg_2/3, and absolute average phase difference (AbsPhaseDiffAvg). AmplAsymAvg and AmplAsymAvg_2/3 are coefficients comparing individual amplitudes of both vocal folds movement in relation to each other for the whole glottal gap and its middle third part, respectively. PhaseAsymAvg and PhaseAsymAvg_2/3 are coefficients comparing the sum of individual amplitudes of vocal fold motion to the amplitude of their resultant movement. Finally, AbsPhaseDiffAvg describes the absolute mean value of phase difference for whole vocal folds. The parameters are precisely described in Supplementary Table 1 and were calculated as described by Sielska-Badurek et al. and Just—inventor of the software (Just et al. 2016; Sielska-Badurek et al. 2019).

First, we used the Shapiro–Wilk test to check assumptions regarding the distribution of analyzed variables. Depending on fulfilled assumptions, for two groups of patients (benign–malignant and involved–healthy), either non-parametric U Mann–Whitney test or parametric T test was performed.

For three groups (norm–benign–malignant), we used a one-way ANOVA test. Depending on fulfilled suppositions, we chose either parametric test to analyze variance or non-parametric Kruskal–Wallis test. If the results of the test showed significant differences between mean values or distribution functions, we performed appropriate post hoc tests: non-parametric multiple comparison test or least significant differences (LSD) test.

Next, we performed a ROC analysis in two variants to check which of the parameters would be a good classifier. To do that, patients were split into two, creating a two-state variable. For the first variant, patients were split into norm and benign + malignant clusters, and for the second variant into benign and malignant clusters. The ROC curve was plotted for each of the parameters and the AUC area was calculated. Results showed which parameters are significant as classifiers. Next, we used Youden index, to determine a cut-off point for each of the parameters. Youden Index is a value giving information on the maximum distance of the points on the ROC curve from the diagonal of the square with sides equal to 1 (red line on the ROC diagrams). For this point, we presented the values of sensitivity and specificity. The parameters’ values were later correlated with the presence and type of lesion, dividing them into either boosters or inhibitors. Values of boosters were increasing along with the risk of disease (either any lesion or malignancy, depending on the variant of ROC analysis as described above), and values of inhibitors were decreasing conversely to the growing risk of disease.

Calculations were performed using STATISTICA.PL software, version 13.3 (Statsoft, Cracov) and Microsoft Excel.

Aside from numerical values of parameters, also four graphs were generated—one for each parameter group and a phonovibrogram visualizing the movement of vocal fold edges in relation to the glottal axis. In Fig. 2, we present a comparison of analysis results for two patients: Subject 1—a 56-year-old male, complaining of mild to severe dysphonia. Initial examination revealed a polypoid mass on the right vocal fold. Subjective kymographic assessment showed a moderate decrease in amplitude of the involved vocal fold, without significant asymmetry or phase difference. As shown on the phonovibrogram, there was a non-opening area of the glottis encompassing the lesion. The average amplitude of the involved vocal fold reached 2.7%FL and the resultant amplitude was 6.5%FL. Values of average amplitude are expressed in the percentage of the total length of the respective vocal fold (percent of fold length–%FL). Surgical removal of the mass was performed, and the patient achieved satisfying postoperative vocal results. In this case, histopathology confirmed the initial diagnosis. Subject 2—a 69-year-old male, complaining of severe dysphonia, initial examination revealed a large polypoid mass on the right vocal fold. Subjective analysis revealed a significant decrease in the amplitude of the involved vocal fold—especially visible on the phonovibrogram as the dark red color of the lower part of the diagram. On phonovibrograms, brightness increases along with the amplitude of oscillations. There was no significant elevation of asymmetry or phase difference. As noticed in the preoperative HSV examination, in this patient, the average amplitude of the involved vocal fold reached 1.0%FL, and the resultant amplitude was 2.5%FL. Decreased (in comparison to Subject 1) values of amplitude indicated a risk of deep tissue infiltration. A surgical approach (TOLMS) was implemented with the removal of the mass along with its pedicle—postoperative histopathology revealed squamous cell carcinoma with no malignant infiltration of the pedicle and free vocal fold edge. This comparison is an accurate example of how advanced visualization techniques with objective analysis can assist clinicians in establishing initial diagnosis. A detailed description of STV analysis for the two subjects is provided in Fig. 2.

All the parameters were compared in three groups—healthy participants, benign lesion subjects, and malignant lesion subjects. One-way ANOVA test was performed to determine if the distribution of the parameters differs between all three groups altogether. Depending on fulfilled suppositions, we chose either parametric test to analyze variance or non-parametric Kruskal–Wallis test (one-way ANOVA on ranks). Fundamental frequency did not differ between the groups. Among STV parameters, significant diversities between the groups were found for 10 out of 14. Detailed results are shown in Table 1.

Next, we performed a bilateral multiple comparison test to find out if the parameters are able to assess subjects between norm, benign, and malignant lesion groups, giving hope to use them as a predictive parameter. In our previous studies (Malinowski et al. 2023), we found out that there are more parameters able to facilitate the detection of the presence of any organic lesion (both benign and malignant), than to facilitate the prediction of malignancy—demonstrating statistically significant differences between benign and malignant organic lesion. Out of the ten STV parameters demonstrating significant differences between all the groups, we determined that four of them also show differences between benign and malignant lesion groups. Those parameters were: AmpAvg, AmpInvolvedAvg, AmpAsymAvg, and AbsPhaseDiffAvg. The values of amplitude were decreasing from benign to malignant lesions, on the other hand, asymmetry and phase difference were increasing with the appearance of malignancy. Results for those parameters are shown in Fig. 3. The remaining six parameters: (AmpAvg_2/3, non-opening, OQAvg_2/3, AmpAsymAvg_2/3, PhaseAsymAvg, PhaseAsymAvg_2/3) presented similar mathematical trends—decreasing values of amplitude and open quotient and increasing values of asymmetry along with the risk of malignancy, however without reaching statistical significance between benign and malignant lesion groups. Significant differences were found only between healthy subjects and whole organic lesion groups (both benign and malignant altogether).

The parameters comparing involved and healthy vocal fold were subjected to another analysis, comparing values of parameters between involved and healthy vocal folds individually among benign and malignant lesions. We found that the amplitude of healthy vocal folds displayed a higher resultant amplitude than folds involved with organic lesions. This difference was significant for both benign and malignant lesions (Fig. 4).

In the next step, we performed AUC ROC analysis in two variants. In the first variant, we assessed the performance of the parameters in detecting the presence of organic lesions. The best performance was observed for AmplAsymAvg, reaching an AUC of 0.822 with 60% sensitivity and 90% specificity (p < 0.0001). Altogether six parameters reached AUC values above 0.7, with varying sensitivity and specificity as shown in Table 2. The highest sensitivity was reached by PhaseAsymAvg with an AUC of 0.721, sensitivity of 88%, and specificity of 54%, p < 0.0001. On the other hand, the highest specificity was reached by OQAvg_2/3 with an AUC of 0.647, sensitivity of 42%, and specificity of 92%. Detailed results are shown in Table 2 and ROC curves are shown in Fig. 5. The parameters can also be assessed as either boosters or inhibitors. Values of boosters were increasing along with the risk of organic lesion and values of inhibitors were decreasing conversely to the growing risk of lesion. The study showed that inhibitors were amplitude and open-quotient-related parameters, meaning that both the amplitude and opening of vocal folds were decreased in malignant lesions. Boosters were asymmetry and phase difference parameters—those properties of the glottal movement were increased in case of malignancy. For each parameter, by means of the Youden Index, a cutting point was calculated. It was a proposed value for the detection of the presence of any organic lesion. Depending on booster or inhibitor properties of the parameter, its values above (for booster) or below (for inhibitor) the cutting point suggest the presence of an organic lesion; on the other hand, values below (for booster) or above (for inhibitor) indicate normophonic patient, respectively. The sensitivity and specificity of the stated cutting point are also shown in the table.

In the second variant of ROC analysis, we compared the performance of the parameters in the preliminary detection of malignancy among organic lesions. AUC values in this variant were generally lower. The four parameters with the largest values were: AmpInvolvedAvg, AbsPhaseDiffAvg, AmplAsymAvg, and AmplAsymAvg_2/3. The greatest value of AUC was reached by AmplAsymAvg: 0.719 with a sensitivity of 65% and specificity of 76.5% (p < 0.0001). The highest sensitivity was achieved by AmpInvolvedAvg with an AUC of 0.7, sensitivity of 72.5%, and specificity of 65.9% (p = 0.0002). On the other hand, the highest specificity was reached by AbsPhaseDiffAvg with an AUC of 0.676, sensitivity of 57.5%, and specificity of 78.8% (p = 0.0009). Detailed results are shown in Table 3 and Fig. 6. As mentioned above, the parameters displayed properties of either boosters or inhibitors with the same types of parameters belonging to each cluster as in the first variant of ROC analysis. Cutting point was calculated for each parameter as described above.