基於卷積神經網路的多層次醫學影像檢索

Multi-hierarchy Medical Image Retrieval Based on Convolutional Neural Network

謝承曄(Hsieh Cheng-Yeh)

指導教授 ： 羅崇銘

國立政治大學/文學院/圖書資訊與檔案學研究所/碩士(2022年)

https://doi.org/10.6814/NCCU202201220

摘要

隨著醫學影像相關工具功能的增加與進步，醫學影像在醫院中廣泛地被使用。為了 有效管理、檢索與利用醫學影像資料庫中的影像，基於內容的醫學影像檢索系統，能協 助使用者尋找所需的資訊，且在醫學教育、臨床輔助診斷與研究領域上被應用。先前研 究利用卷積神經網路(convolutional neural network, CNN)擷取影像特徵，並成功地建 立醫學影像檢索系統，然而過去研究使用的資料量較少，且沒有呈現出醫學影像在臨床 使用時，多種資訊與關聯性的呈現，除此之外，醫學影像中有許多是由一系列的 2D 連 續切片影像組成，系列內影像皆十分相近，而先前研究沒有對此設計處理流程。因此本 研究廣泛地從公開資料庫中搜集不同醫院產生的各式影像資料集，包括超過 10 個國家， 數十個醫院、學校和實驗室的來源，並整理出 14 種成像模式，以及相對應的 40 種不同 器官及 52 類不同疾病的多層次醫學影像資料庫，總共超過 50 萬張。實驗中按照成像模 式、器官和疾病的層次結構進行分類，使用擷取代表性影像的方法來處理大量的資料， 設計分層結構的卷積神經網路(Multi-level CNN)，在階層訓練中調整階層權重及參數 的設計。由訓練完成的模型擷取特徵建立醫學影像檢索系統，檢索結果呈現同一系列不 重複的 2D 切片影像，以提供更多元的檢索資訊。結果顯示擷取代表性影像的方法能夠 減少 50%的訓練時間，同時提高平均檢索精準度 0.01。以此結合 Multi-level CNN 訓練 分層結構的醫學影像資料庫，達到 0.86 的檢索精準度，高於文獻中使用 ResNet152 的 0.71。本研究提出的影像檢索架構能提升大規模醫學影像檢索系統的速度與精準度，以 多層次影像結構呈現，協助使用者有效率地獲取欲查詢的影像資訊。

關鍵字

醫學影像 ； 基於內容的醫學影像檢索 ； 卷積神經網路 ； 多層次醫學影像分類 ； 分層結構的卷積神經網路

並列摘要

With the advancement of medical technology, medical imaging has been widely used in hospitals. To efficiently manage, retrieve and utilize the images in the medical image databases, the content-based medical image retrieval (CBMIR) systems can help users find the required information. CBMIR is widely used in the fields of medical education, clinical aided diagnosis, and research. Previous studies have used convolutional neural network (CNN) to extract image features and successfully build a medical image retrieval system. However, the amount of data used in previous studies is relatively small, and the presentation of various information and correlations in medical images has not been presented. In addition, many medical images consist of a series of 2D serial slices, which are very similar, and the processing flow has not been established in previous studies. Therefore, this study extensively collected various image datasets generated by different hospitals from public databases, including more than 10 countries, and dozens of sources of hospitals, schools, and laboratories. The dataset has a total of more than 500,000 images, including 14 imaging modalities, 40 organs, and 52 diseases, and experimental data are categorized by imaging modality, organ, and disease level. This study used 2 methods of capturing representative images to process large amounts of data. This study proposes multi-level convolutional neural network (Multi-level CNN) and adjusts layer weights and parameters during the training session. CBMIR system is established by extracting features from the trained model, and the retrieval results present the same series of non-repetitive 2D slice images to provide more diverse search information. The experimental results show that the method of capturing representative images can reduce the training time by 50% and improve the average retrieval accuracy by 0.01. Multi-level CNN combined with representative image methods achieves a retrieval accuracy of 0.86, which is higher than 0.71 using ResNet152 in the literature. The proposed image retrieval architecture can improve the speed and accuracy of large-scale medical image retrieval systems, which are presented in a multi-level image structure to help users efficiently obtain the desired image information.

並列關鍵字

Medical image ； Content-based medical image retrieval ； Convolutional neural network ； Multi-hierarchy medical image ； Multi-level CNN

參考文獻

衛生福利部中央健康保險署. (2019). 全民健保資料庫應用服務. 衛生福利部 Retrieved from https://www.ey.gov.tw/File/536638F3EB20C262?A=C

衛生福利部中央健康保險署. (2020). 健保資料人工智慧應用研討會今盛大舉行，展現AI科技應用成果. Retrieved from https://www.nhi.gov.tw/News_Content.aspx?n=FC05EB85BD57C709&sms=587F1A3D9A03E2AD&s=A250E4E0D3A89837

Ahmad, J., Muhammad, K., Lee, M. Y., & Baik, S. W. (2017). Endoscopic image classification and retrieval using clustered convolutional features. Journal of medical systems, 41(12), 1-12.

Akin, O., Elnajjar, P., Heller, M., Jarosz, R., Erickson, B., Kirk, S., & Filippini, J. (2016). Radiology data from the cancer genome atlas kidney renal clear cell carcinoma [TCGA-KIRC] collection. The Cancer Imaging Archive.

Al-Dhabyani, W., Gomaa, M., Khaled, H., & Fahmy, A. (2020). Dataset of breast ultrasound images. Data in brief, 28, 104863.