2026-11-17 ロイヤルメルボルン工科大学(RMIT)
<関連情報>
- https://www.rmit.edu.au/news/all-news/2026/nov/thermul-study
- https://journals.sagepub.com/doi/10.1177/19322968261501979
糖尿病関連足潰瘍の初回診察時における治癒予測の自動化(Automated Prediction of Diabetes-Related Foot Ulcer Healing at Initial Assessment)
Nila Novita Sari, Quoc Cuong Ngo, Rajna Ogrin, Samantha Hanna, Kate Waller, Barbara Polus, Richard J MacIsaac, Elif Ekinci, Dinesh Kant Kumar
Journal of Diabetes Science and Technology(Research Article)
First published online: 2026-11-08
DOI: 10.1177/19322968261501979
糖尿病関連足潰瘍(DFU)は、治癒が遅れると感染や下肢切断につながる重大な合併症である。従来は、初診後4週間の創傷面積の縮小率などを用いて治癒の遅れを評価するため、初診時点での予測が困難だった。本研究では、赤外線サーモグラフィ画像を解析して治癒可能性を予測するAIシステム「TherMUL」を開発・評価した。オーストラリアの専門足病診療施設2か所で、糖尿病性末梢神経障害を伴う患者の創傷画像を取得し、濃度共起行列(GLCM)による画像テクスチャ解析とサポートベクターマシン(SVM)を組み合わせて、12週間後の治癒状態を予測した。モデル開発には56人の68か所の潰瘍を使用し、平均AUCは0.635、未治癒潰瘍の検出感度は75.8%、治癒潰瘍の特異度は53%だった。さらに、独立した7症例での予備評価では5症例の予測が的中した。本技術は、初診時の非接触画像から治癒遅延リスクを早期に把握し、重点的な治療やフォローアップにつなげる可能性がある。一方、予測精度はまだ限定的であり、対象集団の拡大や異なる診療環境での外部検証が必要である。
Diabetes-related foot ulcers (DFUs) are serious complications that can lead to infection and lower-limb amputation when healing is delayed. Conventional assessment often relies on changes in wound area over the first four weeks of treatment, making early prediction at the initial clinical presentation difficult. This study developed and evaluated TherMUL, an AI-enabled smartphone-based thermal imaging system designed to predict DFU healing outcomes. Wound images were collected from patients with diabetic peripheral neuropathy at two specialist foot clinics in Australia. The system combined grey-level co-occurrence matrix (GLCM) texture analysis with a support vector machine (SVM) classifier to predict healing status at 12 weeks. The model was developed using 68 ulcers from 56 participants. It achieved a mean area under the curve (AUC) of 0.635, with 75.8% sensitivity for detecting non-healing ulcers and 53% specificity for identifying healed ulcers. In a preliminary assessment using an independent dataset of seven cases, five predictions were correct. This non-contact approach could help clinicians identify ulcers at risk of delayed healing during the initial consultation and support earlier targeted treatment and follow-up. However, its predictive performance remains moderate, and further validation using larger and more diverse patient populations and clinical settings is required.