Explainable deep learning localizes age-discriminative candidate regions on hand radiographs of adult women.
Authors
Affiliations (10)
Affiliations (10)
- School of Biomedical Engineering, Korea University, Seoul, Republic of Korea.
- School of Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Republic of Korea.
- School of Electrical Engineering and Computer Science, Gwangju Institute of Science and Technology (GIST), Gwangju, Republic of Korea.
- Department of Mathematical Sciences, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Republic of Korea.
- Department of Orthopaedic Surgery, Chungnam National University Sejong Hospital, Chungnam National University College of Medicine, 20 Bodeum 7-Ro, Sejong, 30099, Republic of Korea.
- Department of Orthopaedic Surgery, Chungnam National University Sejong Hospital, Chungnam National University College of Medicine, 20 Bodeum 7-Ro, Sejong, 30099, Republic of Korea. [email protected].
- Department of Orthopaedic Surgery, Chungnam National University Hospital, Chungnam National University College of Medicine, 282 Munhwa-Ro, Jung-Gu, Daejeon, 35015, Republic of Korea.
- Department of Orthopaedic Surgery, Chungnam National University Hospital, Chungnam National University College of Medicine, 282 Munhwa-Ro, Jung-Gu, Daejeon, 35015, Republic of Korea. [email protected].
- Department of Radiology, Seoul National University Hospital, Seoul National University College of Medicine, Seoul, Republic of Korea.
- Kim Jaechul Graduate School of AI, Korea Advanced Institute of Science and Technology (KAIST), Seoul, 06764, Republic of Korea.
Abstract
To determine whether deep learning can distinguish younger from older adult women on hand radiographs and to identify candidate anatomical regions for age-group discrimination using explainable artificial intelligence and quantitative cortical measurements. This retrospective, single-center, exploratory study included 2,221 women sampled consecutively within each age group during 2020-2023 (1,134 aged 20-45 and 1,087 aged 55-80 years at radiography). Patients contributed one posteroanterior radiograph and were assigned to mutually exclusive training, validation, or test sets. Five architectures were evaluated on the same 334 held-out test radiographs (171 younger; 163 older), with performance averaged across five random-seed runs. DenseNet121 was selected by descriptive comparison for gradient-weighted class activation mapping (Grad-CAM) and ray-based analyses with retraining. Manual cortical measurements of test radiographs yielded metacarpal and proximal phalangeal indices for candidate rays identified by Grad-CAM. DenseNet121 achieved the highest mean area under the receiver operating characteristic curve (AUC, 0.996; 95% CI, 0.992-0.999). Grad-CAM showed prominent activation around the third proximal phalanx, with additional activation in the second and fourth rays. Performance remained high after ray masking; the third-ray-only model achieved the highest mean accuracy among single-ray models. The third proximal phalangeal index had the highest AUC among the six indices (0.956; 95% CI, 0.937-0.976). Grad-CAM, ray-based analyses, and quantitative cortical measurements support the third proximal phalanx as an anatomically interpretable candidate region for age-group discrimination in adult women. Integrating explainable artificial intelligence with quantitative radiographic assessment provides an exploratory workflow for investigating hand skeletal aging, with potential relevance to broader musculoskeletal imaging research.