Early detection of laryngeal squamous cell carcinoma (SCC) is critical for improving prognosis and preserving laryngeal function. In routine clinical practice, outpatient endoscopic examinations performed using white light (WL) and narrow-band imaging (NBI) are widely used to detect vocal fold lesions, but visual examinations can be challenging due to subtle lesion appearance or variability in imaging conditions. While deep learning-based SCC detection methods have shown promising results, their training often depends on large labeled datasets that are not always available due to annotation constraints. Consequently, current approaches do not fully exploit information encoded in large, routinely acquired unlabeled data. In this work, we propose a semi-supervised learning framework for laryngeal SCC detection in endoscopic frames that integrates contrastive representation learning into a single-stage object detection architecture. During training, unlabeled images are processed through a contrastive branch operating on informative patch-level representations extracted from the detection network. By enforcing representation consistency across different augmented views of relevant tissue regions, the proposed strategy encourages the learning of more robust and lesion-aware features under heterogeneous imaging conditions, while preserving a standard inference pipeline without additional supervision. We evaluated the proposed framework on three datasets collected in two outpatient clinics, comprising 778 endoscopic frames from 79 patients, enabling assessment across different acquisition settings and clinical cohorts. Experimental results demonstrated consistent improvements over supervised training across all evaluation datasets and imaging modalities. In particular, compared with standard supervised training, the proposed framework improved [email protected] improvements of ＋5.80, ＋4.15, and ＋8.18 points across the three evaluation datasets while also improving Recall across cohorts. These findings suggest that the proposed semi-supervised framework can enhance robustness and generalization in real-world laryngoscopic SCC detection, particularly under heterogeneous acquisition conditions and domain shift.
Dual-branch semi-supervised patch learning for laryngeal squamous cell carcinoma detection in endoscopic images
Sampieri, Claudio;Mora, Francesco;Peretti, Giorgio;Bellini, Elisa;
2026-01-01
Abstract
Early detection of laryngeal squamous cell carcinoma (SCC) is critical for improving prognosis and preserving laryngeal function. In routine clinical practice, outpatient endoscopic examinations performed using white light (WL) and narrow-band imaging (NBI) are widely used to detect vocal fold lesions, but visual examinations can be challenging due to subtle lesion appearance or variability in imaging conditions. While deep learning-based SCC detection methods have shown promising results, their training often depends on large labeled datasets that are not always available due to annotation constraints. Consequently, current approaches do not fully exploit information encoded in large, routinely acquired unlabeled data. In this work, we propose a semi-supervised learning framework for laryngeal SCC detection in endoscopic frames that integrates contrastive representation learning into a single-stage object detection architecture. During training, unlabeled images are processed through a contrastive branch operating on informative patch-level representations extracted from the detection network. By enforcing representation consistency across different augmented views of relevant tissue regions, the proposed strategy encourages the learning of more robust and lesion-aware features under heterogeneous imaging conditions, while preserving a standard inference pipeline without additional supervision. We evaluated the proposed framework on three datasets collected in two outpatient clinics, comprising 778 endoscopic frames from 79 patients, enabling assessment across different acquisition settings and clinical cohorts. Experimental results demonstrated consistent improvements over supervised training across all evaluation datasets and imaging modalities. In particular, compared with standard supervised training, the proposed framework improved [email protected] improvements of ＋5.80, ＋4.15, and ＋8.18 points across the three evaluation datasets while also improving Recall across cohorts. These findings suggest that the proposed semi-supervised framework can enhance robustness and generalization in real-world laryngoscopic SCC detection, particularly under heterogeneous acquisition conditions and domain shift.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
