With recent advances in artificial intelligence (AI) and machine learning (ML) specifically, computer vision has improved performance in classification, segmentation, and object detection tasks in medical imaging and can now assist clinicians in clinical decision-making.^1,2 Computer vision has been widely applied to video-endoscopy of the upper airway tract (UAT) to detect, classify, and segment pathologies in the nasopharynx, oropharynx, and larynx.^3–9 It holds the promise to reduce interrater variability in interpretation of UAT video-endoscopy and enhance reliability across varied clinical settings, thus improving collective diagnostic performance.

Object detection is a computer vision technique that helps identify and localize anatomical regions or pathology. Object detection consists of two specific tasks: object localization and object classification.^10–12 Object localization identifies specific objects through bounding boxes or other markers, while object classification bins the objects into a predefined class or category, such as benign or malignant lesion.^10–12 One of the earliest techniques used for object detection was the sliding window technique. This was followed by higher performing methods involving Convolution Neural Networks (CNNs) like R-CNN, Fast R-CNN, and Faster CNN. ^13–15 You Only Look Once (YOLO) was developed in 2015 and demonstrated better real-time performance than all previous algorithms in object detection. ¹⁶ YOLO enables real-time object detection by performing classification and localization in a single forward pass, a critical feature for video applications.¹⁶

Interest and research in YOLO-based systems has been rapidly growing in gastrointestinal endoscopy, particularly for real-time detection of colorectal polyps. Numerous studies have demonstrated its effectiveness, with models achieving detection recall rate as high as 98%.^17–21 These models have proven effective not only in detecting large polyps but also in identifying small, flat, and difficult-to-visualize lesions that may be missed with routine evaluation.²² Systematic and scoping reviews of AI in colonoscopy have highlighted YOLO as an effective architecture for polyp detection, particularly where rapid inference and deployment are critical.^23–25 These reviews also report trends toward improved adenoma detection rates in YOLO-based systems compared with other computer aided detection (CADe) platforms. Several FDA approved CADe algorithms for colorectal polyp detection are currently in commercial use through systems such as GI genius (Medtronic), SKOUT (Iterative Scopes), MAGENTIQ-COLO (MAGNETIQ EYE) and CADDIE (Olympus), though their reliance on YOLO is not confirmed.

To improve object detection consistency and detection across frames with YOLO, temporal tracking has been implemented in some colorectal studies. Temporal tracking refers to the process of following objects across consecutive video frames by leveraging motion and spatial continuity. In colonoscopy, spatio-temporal models have improved the identification of small polyps by analyzing their appearance and movement across frames.²⁶ Similarly, software tools such as Polyp-Alert made use of tracking to assist endoscopists in real-time colonoscopies for improved polyp detection.²⁷

Given the success of YOLO in colorectal polyp detection studies, its application to other forms of video-based medical imaging, such as stroboscopy and laryngoscopy, presents a promising avenue for real-time lesion detection and clinical decision support. Several teams have notably explored the potential of YOLO in laryngeal cancer and benign vocal fold lesion detection.^28–30 The latest YOLO model used in the laryngology literature is YOLOv8n.²⁸ In this study, we explore the latest versions of YOLO (YOLO11 and YOLO12) for detection and localization of laryngeal polyps in video-stroboscopy frames. The task focuses on identifying the presence of polyps on the vocal folds, rather than diagnosing or classifying amongst different pathological subtypes. To enhance the detection of smaller or less distinct laryngeal polyps, we also implemented a temporal tracking module on top of the YOLO detection models. To our knowledge, this is the first study to integrate YOLO models with temporal tracking for the detection of laryngeal lesions across frames in video-laryngoscopy.

Demographic details are summarized in Table 1, including age, sex and smoking history.

We demonstrated the feasibility of near real-time laryngeal polyp detection in video-stroboscopy using deep learning with the latest YOLO models. YOLO12 outperformed YOLO11 across all metrics, with an F1 score of 68.0% and mAP@0.5 of 70.4%. Furthermore, differences in the tuned tracking parameters, particularly the frame similarity thresholds, reflected that, while both models had missed detections during motion or angle changes, YOLO11 was less stable and necessitated stricter temporal matching to maintain tracking reliability.

YOLO12’s higher performance is consistent with its architectural advancements, which includes the incorporation of enhanced attention mechanisms and improved feature aggregation capabilities.^34,37 The inclusion of the temporal tracking module modestly improved performance with slight increases in the F1 score and the mAP metrics. Prior research in endoscopic video analysis has noted that single-frame detectors can be susceptible to momentary misdetections and cannot leverage temporal context.³⁸ YOLO11’s bigger leap in performance with temporal tracking compared to YOLO12 may be explained by a lower baseline performance with more room for improvement compared to YOLO12, with its stronger base performance with less room for relative gain.

Grad-CAM has been employed by previous researchers to explain the performance of a model.^39,40 In our case, the visualizations reinforced the observed performance differences between YOLO12 and YOLO11. YOLO12 showed more concentrated activation near lesions, whereas YOLO11 displayed scattered attention. (Figures 2, 3). These shortcomings may arise from differences in model architecture. However, it is important to interpret these results cautiously, as Grad-CAM provides only a coarse approximation of model attention and may not always reflect the true reasoning behind a prediction.⁴¹

Despite these promising results, YOLO performance in vocal fold polyp detection remained behind reported performances in polyp detection for colonoscopy. Recent colonoscopy models have achieved precision and recall exceeding 90%, with F1 scores and mAP often above 85%.^42,43 For example, a YOLOv8-based system recently reported a precision of 95.6% and an F1 score of 92.4% in real-time colorectal polyp detection.⁴³ Colonoscopy offers a more favorable landscape for lesion detection. The colon provides a relatively uniform tubular anatomy with consistent lighting, fewer obstructions, and slower flexible endoscope movements facilitated by patient sedation. Polyps typically present as protruding lesions against a homogenous background and persist across multiple frames, enabling better learning of visual features.⁴⁴

In contrast to the colon, the UAT anatomy, specifically that of the larynx, is structurally complex. The vocal folds, epiglottis, and surrounding tissues produce variable lighting, shadowing, and movement artifacts in the awake patient. The rapid laryngeal movements during phonation, breath, cough, gagging, or swallowing create motion blur and reduced visibility of subtle vocal fold lesions, such as polyps. Frames may contain occlusions, inconsistent viewpoints, and fewer clear, repeated views of the same lesion, limiting opportunities for the model to detect a polyp across frames. These factors make video-laryngoscopy a more challenging task for polyp detection compared to colonoscopy.

In colonoscopy, AI-assisted systems have improved adenoma detection rates by offering a second observer effect, prompting endoscopists to re-examine areas flagged by AI.⁴⁵ A similar paradigm could benefit laryngoscopy, where vocal fold polyps and other vocal fold lesions may be subtle and go undetected by the otolaryngologist. An object detection AI algorithm, such as the one we created with YOLO, could flag polyps, potentially improving diagnostic yield during routine examinations, and assisting in-office procedural interventions such as laser ablation.

Several limitations of this study must be acknowledged. The dataset, while relatively large in frame count, comprised only 55 patients, limiting the diversity of polyp presentations and anatomical variability. The tracking module used was effective but rudimentary, relying on ORB-based visual similarity and template matching. While it helped improve recall, it may propagate false positives and lacks the accuracy and precision of deep learning models. The moderate F1 score indicates that some detections are still missed or incorrectly flagged, limiting consideration for immediate clinical deployment. Our pipeline detects lesions at 21.4 fps compared to the 30 fps rate of clinical laryngoscopic and stroboscopic examinations. While suitable for near real-time applications, it does not yet achieve true real-time performance. Achieving full real-time integration within a stroboscopic workflow will require additional components such as strobe light synchronization, synthetic cycle reconstruction, and live video augmentation, all of which represent important directions for future development.

Future work will focus on expanding the dataset to include a larger and more diverse cohort of patients with varied laryngeal pathologies. Cross-institutional collaborations will be essential for this effort, as they facilitate access to heterogeneous patient populations, diverse clinical practices, and a range of imaging hardware. Such collaborations are critical for enhancing model generalizability across demographics, healthcare settings, and for validating performance on diverse cohorts and equipment types. Advanced temporal modeling, such as the use of three-dimensional convolutional neural networks or transformer-based architectures, may further enhance detection accuracy by integrating optical flow, motion, and contextual information across frames. Developing multi-class detection frameworks capable of distinguishing between multiple lesion types, such as nodules, cysts, and malignancies, would more accurately reflect the clinical heterogeneity observed in laryngoscopy, as previously explored by Baldini et al.²⁸ Making use of Grad-CAM alternatives such as HiResCAM could also provide better insights into model interpretation, allowing for more informed training adjustments. Ultimately, these advancements could pave the way for accurate, real-time, AI-assisted diagnostic tools to support otolaryngologists in the early detection and management of vocal fold lesions.

This study presents the first approach to near real-time deep learning-based laryngeal benign lesion detection using YOLO-based object detection with a temporal tracking system. YOLO12, enhanced with a simple tracking pipeline, demonstrated good accuracy in identifying and localizing vocal fold polyps in stroboscopic videos, demonstrating potential for AI-assisted diagnosis. However, performance lagged that reported in colonoscopy, likely due to the anatomical complexity of the larynx and combined motion of the flexible laryngoscope and the vocal folds. With larger and more diverse datasets, advanced temporal modeling, and multi-class detection capabilities, future AI systems could assist otolaryngologists with the detection and identification of varied vocal fold pathologies in real-time, reducing the likelihood of missed lesions amongst less experienced clinicians.