前瞻研究成果-1 – 人工智慧技術暨全幅健康照護聯合研究中心

前瞻研究成果

Researches

▎用AI打造急診全新流程

The AI-assisted Innovation In Patient Flow of Emergency Department

►國立臺灣大學陳銘憲特聘教授 / Distinguished Prof. Ming-Syan Chen, National Taiwan University

自2020 年三月起，臺大AI中心結合臺灣大學AI技術應用與臺大醫院急診醫學部，聯手企業夥伴，執行科技部 AI 拔尖整合計畫。運用人工智慧技術，在臺大醫院急診室打造智慧急診全新流程。藉由導入精準電子檢傷、病史即時分析、胸部 X 光即時危險辨識、安全離部的決策推薦、院內心跳停止事件的高危風險預測、打造物聯網場域建立高危病患預警，與醫護過勞預測等各種 AI 模型，提供醫護即時預測與診療輔助建議，部分流程更有效縮短檢驗診療時間從數小時到數十分鐘，大幅縮短病患停留急診期間，有效降低誤診率與優化醫療品質，最終解決急診壅塞問題。

Since March 2020, Most Joint Research Center for AI Technology and All Vista Healthcare combined AI technology applications of NTU, the emergency department of NTUH, and industrial partners to execute Capstone project. With AI technology, the Capstone project team is establishing the brand-new intelligent patient flow of emergency departments by deploying various categories of AI models.It includes precise triage, real-time medical history analysis, real-time X-Ray recognition for chest inappropriate intubation, appropriate discharge recommendation, high-risk prediction for In-Hospital Cardiac Arrest, proactive warning for patient’s critical illness, and medical personnel overwork prediction.

We expect to provide real-time prediction and treatment recommendations for assisting the doctor’s diagnosis process. Furthermore, some diagnosis processes could be shortened substantially from hours to minutes. Our goal is not only to reduce the erroneous diagnosis effectively but to optimize medical quality. Finally, it could resolve the ED crowding phenomenon significantly.

Reference :

▲影像辨識疼痛指數 | Image recognition application on pain rating index

▲AI輔助的創新急診流程 | An AI assisted Innovated Emergency Medicine Patient Flow

此研究歸屬科技部 AI 專案計畫執行成果,詳細資訊請參考附錄之計畫總表第 1 項
For the name of the project which output this research, please refer to project serial no. 1 on the List of MOST AI projects on Appendix

▎金融意見探勘與應用

Fine-Grained Financial Opinion Mining

► 國立臺灣大學陳信希特聘教授 / Distinguished Prof. Hsin-Hsi Chen, National Taiwan University

本團隊持續提出金融意見探勘領域的創新研究議題，透過建立代表性資料集，於國際研討會中引領該領域研究人員之發展方向。於社群媒體投資人意見，本團隊關注於數字理解問題，並透過改良編碼方式突破現有技術上對文本中數字理解的不足；於專業投資人意見，本團隊開啟論證分析的基石，引導相關研究進入更細粒度的方向；本團隊亦點出財務虛假資訊探勘的重要方向。研究成果亦落地開發成展示系統，並於金融科技競賽中獲得第一名的佳績。

As one of the dominant teams in financial opinion mining, we have proposed several novel research issues in financial opinion mining and lead the development of this field by offering benchmark datasets. We made up the shortcomings of previous embedding methods in numeral representation, and designed a multi-task architecture for solving the numeral attachment task. The first exploration of argument mining in financial narratives was also present. We also pointed out the important issue in the financial false information detection task.

The demonstration systems based on our results were developed, including visualizing the textual data on financial social media platforms to indicators and an opinion-based personalized recommendation system. Additionally, the proposed system won the first prize in the FinTech Competition.

Reference :

▲ System Flowchart

▲ 金融科技競賽冠軍

此研究歸屬科技部 AI 專案計畫執行成果,詳細資訊請參考附錄之計畫總表第 3 項
For the name of the project which output this research, please refer to project serial no. 3 on the List of MOST AI projects on Appendix

▎基於深度學習之阿茲海默症快篩系統

NLP-based Deep Learning for Fast Screening of Alzheimer’s Disease

► 國立臺灣大學傅立成講座教授 / Chair Prof. Li-Chen Fu, National Taiwan University

我們替長者研發出是否罹患輕微認知退化或阿茲海默症的一個快篩系統，利用神經心理學原理，透過訪談長者的說話內容，輸入至一能處理自然語言的深度學習網路架構，即可迅速就受訪長者的說話語音或轉譯的文字進行分析及分類，以達到早期偵測失智症的目標。本系統經測試於臺大醫院所收的個案中顯示，偵測輕微認知退化或阿茲海默症的準確率分別高達82%及91%，高於國際學術界的相關報告，在今日老齡化的社會中，替長者認知健康提供多一項保障。

The team has designed and implemented an effective screening system, which could analyze whether an elder has Mild Cognitive Impairment (MCI) or Alzheimer’s disease (AD) on the basis of neuropsychology.

Empowered by a tailored Neural Network that can process natural languages, our system can efficiently analyze and classify spoken text or audio of an elder to detect dementia in an early phase. The accuracy of detection on MCI and AD with the test cases from the National Taiwan University Hospital is as high as 82% and 91%, which shows better results than other contemporary works. Hence, as Taiwan has become an aged society, it is necessary to have such system to safeguard the cognitive health of the elders.

Reference :

https://reurl.cc/mL4j99

▲我們會給長者做神經心理學的測試，接著使用深度學習的技術對長者說出來的文字分析，判斷長者是否可能患有阿茲海默症
The elder will undergo the neuropsychology tests. Then, a deep-learning-based system will analyze the spoken text of the elder and show whether the elder might have Alzheimer’s disease.

此研究歸屬科技部 AI 專案計畫執行成果,詳細資訊請參考附錄之計畫總表第 23 項
For the name of the project which output this research, please refer to project serial no. 23 on the List of MOST AI projects on Appendix

▎YOLOv4 and Scaled-YOLOv4

► 中央研究院廖鴻源博士 / Dr. Mark Liao Hong-yuan, Academia Sinica

YOLOv4 是中研院 AI 團隊與義隆電子一個合作案的產出。它在2020年9月28日到2020年12月15日雄踞在以 MS COCO資料集為基礎的物件偵測競賽排行榜中名列世界第一達到兩個半月之久。它的原始碼開放給全世界使用。臺灣包括竹科、南科等高科技廠商都使用此技術為基礎開發它們的產品。中研院團隊也將技術報告放在archive網站上面成為技術報告,截至2021年9月7日止,該技術報告已被引用 1498 次之多(Google Scholar)。

YOLO4 is the output of a cooperation project between Academia Sinica and ELAN electronics Inc. From September 28, 2020 to December 15, 2020, it ranked first in the world in the object detection competition based on the MS COCO dataset. It defeated all world-class companies, including Google, Facebook, Microsoft, Qualcomm, Amazon, etc. We open the source codes of YOLOv4 to the whole world.

Many high-tech manufacturers in Taiwan, including Hsinchu and Southern Science Park, use this technology as a basis to develop their products. The Academia Sinica research team also turned the report into an archive technical report. As of September 7, 2021, this technical report has been cited 1498 times (Google Scholar).

Reference :

◀ 用YOLOv4執行夜間車輛偵測的結果
Using YOLOv4 to perform vehicle detection during the night.

◀ 用YOLOv4執行夜間車輛偵測的結果
Using YOLOv4 to perform vehicle detection during the day time.

此研究歸屬科技部 AI 專案計畫執行成果,詳細資訊請參考附錄之計畫總表第 8 項
For the name of the project which output this research, please refer to project serial no. 8 on the List of MOST AI projects on Appendix

▎三維點雲模型視覺分析

3D Point Cloud Visual Analysis

► 國立臺灣大學王鈺強教授 / Prof. Yu-Chiang Frank Wang, National Taiwan University

臺灣大學的視覺與學習實驗室提出三維圖卷積網路，結合深度學習的訓練方式，能夠有效辨識立體資訊(點雲)當中的物體類別與位置。該方法相較其他團隊提出的更為有效率，能在物體任意平移、縮放下，仍保持不變的準確率，在實際應用中獲得更好的表現。該演算法可以應用於自駕車、無人機、機器人等系統，更加準確地辨識行人與周遭物體或場景關係，降低意外發生機率，提昇乘客或使用者安全性。該研究成果於2020年發表在國際頂尖的電腦視覺會議CVPR。

3D Graph Convolution Network is proposed by Vision and Learning Lab at National Taiwan University. Trained with deep learning algorithms, the model is able to recognize various object categories and position in 3D data (i.e. point clouds).

3D-GCN is more efficient than previous works, and maintains constant accuracy under severe object translation and scaling, which is desirable in real-world applications. The proposed algorithm can be applied to autonomous driving and related robotics systems, detecting accurate position of pedestrians and the surrounding objects, which would decrease the accident risks and increase the safety of the passengers or users. This research was accepted by top international conference CVPR, 2020.

Reference :

https://reurl.cc/MApyMm

◀受到影像處理的啟發，臺灣大學提出辨識立體資訊的演算法
Motivated by image processing, NTU proposed an algorithm for recognizing 3D patterns.

◀三維圖卷積網路能夠在立體資訊上完成各種任務，包含分類和語意分割
3D Graph Convolution Network is able to perform various tasks on 3D data, including classification and semantic segmentation.

此研究歸屬科技部 AI 專案計畫執行成果,詳細資訊請參考附錄之計畫總表第 7 項
For the name of the project which output this research, please refer to project serial no. 7 on the List of MOST AI projects on Appendix

前瞻研究成果-1 前瞻研究成果-2 前瞻研究成果-3 前瞻研究成果-4 前瞻研究成果-5 前瞻研究成果-6