Agreement analysis between examiner and artificial intelligence-derived measurements of condylar angles on cone-beam computed tomography scans using a semi-automated U-Net workflow
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
Accurate assessment of condylar angulation in temporomandibular joint (TMJ) analysis is essential for detecting morphological abnormalities. Manual evaluation on cone-beam computed tomography (CBCT) scans is time-consuming and examiner dependent, highlighting the potential role of artificial intelligence (AI). This study investigated the agreement between examiner and AI-derived condylar angle measurements using a semi-automated workflow. CBCT scans from 20 patients (40 condyles) were retrospectively analyzed. A total of 902 axial images from 10 patients were used for U-Net model training and validation, while 125 images from 2 patients were reserved for independent testing. Reliability and agreement analyses were then performed on the full dataset comprising 20 patients. The U-Net model achieved strong segmentation performance (Dice coefficient=0.891±0.056; intersection over union=0.807±0.073). Intra-examiner reliability was high (intraclass correlation coefficient=0.804). AI-derived condylar angles demonstrated moderate agreement with manual measurements. Bland–Altman analysis revealed a mean bias of 1.08°, with 74.1% of AI-derived values within ±5° of examiner measurements. Correlation analysis confirmed strong linear and rank associations, though systematic biases were observed. These findings support the feasibility of examiner–AI collaborative workflows to reduce subjectivity and improve reproducibility in TMJ assessment. They also emphasize the importance of reference plane standardization and multiinstitutional validation to establish robust clinical applicability.
Abstract
Accurate assessment of condylar angulation in temporomandibular joint (TMJ) analysis is essential for detecting morphological abnormalities. Manual evaluation on cone-beam computed tomography (CBCT) scans is time-consuming and examiner dependent, highlighting the potential role of artificial intelligence (AI). This study investigated the agreement between examiner and AI-derived condylar angle measurements using a semi-automated workflow. CBCT scans from 20 patients (40 condyles) were retrospectively analyzed. A total of 902 axial images from 10 patients were used for U-Net model training and validation, while 125 images from 2 patients were reserved for independent testing. Reliability and agreement analyses were then performed on the full dataset comprising 20 patients. The U-Net model achieved strong segmentation performance (Dice coefficient=0.891±0.056; intersection over union=0.807±0.073). Intra-examiner reliability was high (intraclass correlation coefficient=0.804). AI-derived condylar angles demonstrated moderate agreement with manual measurements. Bland–Altman analysis revealed a mean bias of 1.08°, with 74.1% of AI-derived values within ±5° of examiner measurements. Correlation analysis confirmed strong linear and rank associations, though systematic biases were observed. These findings support the feasibility of examiner–AI collaborative workflows to reduce subjectivity and improve reproducibility in TMJ assessment. They also emphasize the importance of reference plane standardization and multiinstitutional validation to establish robust clinical applicability.