使用類神經網路的錠劑藥物影像檢索

Pill Image Retrieval using Neural Networks

隨著民眾教育水準的提升以及一些重大醫療用藥疏失的發生，使民眾對於藥物資訊與用藥安全的需求與日劇增。但不論是使用網際網路或是藉由書籍來尋找藥物資訊，通常只能透過藥名或其他描述藥物功能的關鍵字去查詢。影像資料往往很難用文字去描述，本研究提出一個以內容為基礎的影像檢索方法(Content-Based Image Retrieval, CBIR)，透過錠劑藥物數位影像中形狀、大小、顏色等特徵之擷取，結合類神經網路的分類功能，建立一個錠劑藥物外觀影像檢索模型，提供非文字型的藥物查詢方法。運用數位相機或攝影機取得的藥物影像，先經自動化方式擷取影像內容的特徵，再進行辨識。本研究目前得到的辨識正確率達94％以上，證明了使用此模型作為藥物外觀影像辨識工具的可行性。

With the elevation of peoples' educational level and the occurrence of some serious medication mistakes, the need for correct pharmaceutical information and adequate knowledge of medication safety augments. People often look for medicine information on the Internet or in books. However, the queries are usually executed in the text form using drug names or key words of medicine functions. Images are often difficult to be described with text. A content-based image retrieval (CBIR) method was proposed in this article. Shape, scale, and color features of pill images were extracted first and then fed into neural networks for classification. The pill image retrieval model was built by deploying appropriate features and feed forward neural networks to provide non-text query method. After obtaining pill images via digital camera, features that represented the images were extracted automatically. The features were then processed for recognition. The system has achieved a recognition rate higher than 94%, it proves that a pill image recognition system using this model is feasible.