AI in the Diagnosis of Cholesteatoma: A Systematic Review of Current Evidence

CLINICAL QUESTION

How accurately can artificial intelligence (AI) models diagnose and stage cholesteatoma using CT and otoscopic imaging?

BOTTOM LINE

AI-based models demonstrated high diagnostic accuracy for cholesteatoma detection and staging, particularly when using convolutional neural network (CNN) architectures and CT imaging. Although most studies showed excellent internal validation performance, limited external validation and predominantly retrospective designs currently restrict routine clinical implementation.

BACKGROUND: Cholesteatoma is a destructive middle ear lesion associated with hearing loss, bone erosion, vestibular dysfunction, and intracranial complications if untreated. While high-resolution CT provides anatomic detail, distinguishing cholesteatoma from chronic inflammatory middle ear disease remains challenging. Recent advances in AI and deep learning have prompted increasing interest in automated imaging-based diagnosis and surgical planning support.

STUDY DESIGN: Systematic review conducted according to PRISMA guidelines. PubMed, Scopus, and Web of Science were searched through September 2025 for studies evaluating AI, machine learning, or deep learning models for cholesteatoma diagnosis using CT, MRI, or otoscopic imaging. Seven studies met the inclusion criteria.

SETTING: Multinational studies from Turkey, Japan, China, Morocco, and the U.S. involving academic tertiary-care institutions.

SYNOPSIS: Seven studies evaluating AI-based cholesteatoma diagnosis were included. Six studies used temporal bone CT imaging, while one analyzed otoscopic images. Most investigations employed CNN-based deep learning architectures, including DenseNet, MobileNetV2, ResNet50, Inception-V3, and Xception. Internal validation accuracies generally exceeded 90%, particularly in single-center studies.

Among CT-based models, ResNet50 achieved 93.3% accuracy in differentiating chronic otitis media with and without cholesteatoma, while DenseNet201 demonstrated approximately 91% accuracy comparable to diffusion-weighted MRI. A multicenter 3D CNN model incorporating automated region-of-interest detection achieved internal and external accuracies of 87.8% and 84.3%, respectively, while prospectively aiding surgical planning in 90.1% of cases.

Otoscopy-based AI also demonstrated strong diagnostic potential. DenseNet201 achieved 98.5% accuracy and AUROC 0.999 for differentiating cholesteatoma from normal tympanic membranes, although performance declined when distinguishing cholesteatoma from other abnormal middle ear conditions. Several studies additionally showed AI performance comparable to or exceeding human readers for diagnostic and staging tasks. Explainability methods such as Grad-CAM frequently localized clinically relevant anatomic regions, potentially improving clinician trust and interpretability.

The authors emphasize that despite encouraging results, most included studies were retrospective, single-center investigations with substantial risk of bias and limited external validation. Generalizability concerns, potential overfitting, and inconsistent reporting of explainability and calibration remain important barriers to clinical adoption. The review concludes that prospective multicenter validation studies with standardized outcome reporting and clinical impact assessment are needed before widespread implementation of AI-assisted cholesteatoma diagnosis.

CITATION: Shabi SM, et al. Artificial intelligence in the diagnosis of cholesteatoma: a systematic review of current evidence. Cureus. 2025;17:e96154. doi:10.7759/cureus.96154.