Deep learning techniques for otologic imaging: a systematic review of segmentation methods

Developing segmentation models for otologic structures in medical images is fundamental for automated analysis and supporting clinical decision-making in ear-related patient care and surgical management. However, this task remains highly challenging because the ear contains small and morphologically complex substructures that often exhibit low visual contrast against surrounding tissues, making manual annotation labor-intensive and prone to inconsistency. Recent advances in deep learning (DL) have transformed medical image analysis across multiple clinical specialties, presenting new architectures to overcome these challenges. Beyond conventional convolutional neural networks (CNN), state-of-the-art approaches include vision transformers, diffusion models, and large-scale foundation models capable of capturing long-range dependencies, modelling sub-millimeter anatomy, and generalizing across variable imaging conditions. This systematic review provides the most comprehensive evaluation to date of how these DL models have been applied to otologic segmentation. Although DL has been widely reviewed in broader medical imaging, dedicated coverage of otologic applications is limited. Existing otologic reviews are typically narrow in scope, often confined to the middle ear, and overlook many clinically relevant anatomical domains. They also primarily emphasize CNN-based architectures and do not incorporate recent advances from 2023 to 2025, including transformer-based, diffusion-based, and hybrid architectures that integrate domain-prior strategies for enhanced contextual modelling. To address this gap, our review synthesizes research across all major otologic substructures and traces the evolution of DL architectures from CNN-based to hybrid and emerging models. We identify key performance trends and highlight persistent barriers to clinical translation, including severe class imbalance, limited multi-center datasets, variability in imaging, heterogeneous annotation standards, and high computational cost that restrict deployment in hospital settings. Based on these insights, we propose forward-looking research priorities to accelerate translation into clinical workflows. These include explainable and trustworthy DL models to support clinician confidence and regulatory approval, lightweight architectures suitable for diverse resource environments, domain adaptation across scanners and institutions, and scalable deployment frameworks compatible with diagnostic and surgical navigation systems. By integrating current advances, limitations, and future opportunities, this review presents a structured roadmap for developing clinically deployable DL models for otologic segmentation and outlines how these technologies can enhance future diagnostic and surgical practice.