| 摘要: | 背景：糖尿病是一種與飲食相關的慢性疾病，碳水化合物的攝取量與種類選擇，在調控血糖上扮演著一個很重要的角色。其中，升糖指數(GI)是根據食物對於血糖的影響，進而將食物分成高、中、低三類的一種概念。此外，人工智慧(AI)用於辨識影像的方法也逐漸成熟，對於應用在飲食評估上潛在著前瞻性與可行性。目前臺灣尚未建置臺灣常見食物影像資料庫，且無應用AI在糖尿病營養照護的研究。
目的：本研究目的主要是訓練AI辨識台灣常見的升糖指數食物影像，並以糖尿病病人所拍攝紀錄的食物影像日誌來驗證AI辨識升糖指數食物的準確度。
方法：在AI食物影像辨識訓練中，我們首先建立了一個包含150種台灣常見GI食物的食物影像暨營養資料庫，初步挑選三種深度學習模型 （ResNeXt-101-32x8、ViT_B_16 和ResNet152）來測試AI辨識食物影像的準確度，最後，以糖尿病受試者所拍攝的日常飲食影像來驗證模型的準確度。食物影像來源主要分成三種，網路爬蟲、由本實驗室拍攝的標準食物影像以及臨床糖尿病人食物影像，影像資料集分成訓練集：驗證集：測試集＝8：1：1，訓練條件Batch size設為32。 初步招募21位患有糖尿病的台灣成人，使用本實驗室開發的 「Formosa FoodApp」讓受試者拍照記錄日常飲食攝取狀態，營養師再利用「圖像飲食評估分析法」分析病患所攝取的食物熱量及其GI和升糖負荷值(GL值)，共計收集271天食物影像日誌，3003張食物影像。
結果：150種台灣常見GI食物影像資料庫收集37,301張食物影像，95%來自網路爬蟲（35,521張食物影像），2%來自本實驗室拍攝的食物標準照（759張食物影像），3%來自臨床糖尿病病人的飲食日誌（1,121張食物影像）。在AI模型預訓練階段， ResNeXt-101-32x8、ViT_B_16 和ResNet152三種模型平均的整體top-1準確度達到約85.9%，整體 top-5準確度皆超過96%。 對於不同GI食物的辨識上，高、中、低GI食物的整體top-1辨識準確度分別是85.8%、81.7% 和87.9%，雖然整體辨識率高達8成，但有7種食物的平均辨識準確度低於6成。第二階段為臨床驗證，共有78種糖尿病人所攝取的食物影像進行驗證，整體top-1準確度達到80.8%，其中以高GI食物類別的辨識準確度87.4%為最高，其次為低GI食物的76.9%和中GI 60.2%。 ResNeXt-101、ViT_B_16和ResNet152 分別有9、12、8 種食物辨識準確度為0%，後續分析顯示：三種深度學習模型辨識不佳的主因為「人類也難以辨識（人類的準確度低於60%）」、「預訓練的辨識準確率就較低（模型準確度低於60%）」或是「食物種類間相似度高」等因素。
結論：在這項初步研究中，AI深度模型對於辨識臨床糖尿病受試者的食物影像的體準確度達80.8%，其中辨識高GI食物的準確度為最佳。未來目標是持續收集病患飲食資料，並以臨床影像微調AI模型以增進臨床食物影像辨識準確度，並進行食物影像切割，分析個別食物的營養素並計算GL值。
Background: Carbohydrate (CHO) restriction is crucial in managing diabetes. The glycemic index (GI) and glycemic load (GL) are used to assess the effects of CHO on raising blood glucose levels.
Aim: The purpose of this study was to train artificial intelligence (AI) to recognize commonly consumed GI foods in Taiwan, and to validate the pre-trained AI models using GI food images captured by diabetic patients.
Methods: We first build up a GI food image database which contains 150 commonly consumed Taiwanese CHO foods, and pre-trained three deep learning models (ResNeXt-101-32x8, ViT_B_16 and ResNet152) using the established GI image database. The GI food images database was obtained from three sources: publicly available food images downloaded using web-crawler, standardized food images filmed in the laboratory (Lab), and real-life food images captured by diabetic patients. For the clinical validation, 78 food items of real-life food images (n=3,003 images) was recorded using “Formosa FoodApp” by 21 diabetic patients over 271 days. An image-based dietary assessment was carried out by trained nutritionist to estimate the portion size, GI, GL and nutrient intake of patients. Accuracy was presented as percentage (%) of top-1 or top-5 accuracy: top-1 accuracy as the number of model answer correctly divided by total number of predictions, and top-5 accuracy as the number of top 5 highest probability answers that match the correct answer divided by the total number of predictions.
Result: The Taiwanese GI food image database consisted of 150 food items with a total of 37,301 images, of which 95% was derived from web-crawler (n= 35,521 images), 2% was from standardized studio images (n= 759 images), and 3% was real-life food images derived from dietary record of diabetic patients (n= 1,121 images). In the training stage, all three AI models (ResNeXt-101-32x8, ViT_B_16 and ResNet152) achieved >84% top-1 accuracy, and >96% top-5 accuracy. The average top-1 training accuracy for high GI, medium GI and low GI foods were 85.8%, 81.7% and 87.9%, respectively. For the clinical validation, the overall top-1 accuracy was 80.8% for recognizing 78 food items of real-life food images of diabetic patients. The highest top-1 accuracy was the prediction of high GI foods (87.4%), following low GI foods (76.9%) and medium GI (60.2%). However, ResNeXt-101, ViT_B_16, and ResNet152 had 9, 12, and 8 food items with 0% top-1 accuracy, respectively.
Conclusion: In this pilot study, we found that deep learning achieved 80.8% in recognizing the real-life food images of diabetic patients, with high GI foods yield the highest accuracy. The next goal is to fine-tuning deep leaning models based on real-life food images collected from diabetic patients. |
