應用人工智慧於醫療器材設計暨追蹤電路設計於延遲鎖相迴路和靜態隨機存取記憶體

Application of Artificial Intelligence in Medical Device Design and Design of Tracking Circuit on DLL and SRAM

洪碩宏(Shuo-Hong Hung)

指導教授 ： 陳中平

國立臺灣大學/電機資訊學院/電子工程學研究所/博士(2020年)

https://doi.org/10.6342/NTU202001061

摘要

隨著大數據，雲端存取和高效能電腦逐漸普及，人工智慧技術近年來掀起了第三次發展浪潮，其中人工智慧技術在醫療保健領域的發展潛力非常巨大。本論文提出人工智慧應用於腦波判讀和大腸鏡影像判讀。重度憂鬱症被認為是一種傾向慢性，病程易惡化的疾病，具有與其他症狀高度合併風險。一定比例的重度憂鬱症病患在數個抗憂鬱藥物的治療下沒有好轉，而這類型的病患卻有機會被重複性經顱磁刺激或間歇性脈衝式經顱磁刺激所治療。本研究分析了重度憂鬱症患者的腦電圖信號，分別觀察重複性經顱磁刺激或間歇性脈衝式經顱磁刺激治療之重度憂鬱症患者的特徵來預測其抗憂鬱反應。我們使用機器學習用於區分重複性經顱磁刺激和間歇性脈衝式經顱磁刺激的治療有效以及無效者，其驗證準確率分別為 92%和90.9%。結腸鏡檢查為當今預防大腸癌發生的最佳方法，不過因為其為人為操作的技術，品質管理與監測亦為相當重要的一環，盲腸到達率低會導致結腸鏡檢查後大腸癌發生率提高，因此本論文透過深度學習自動化判斷每一次結腸鏡檢查是否到達盲腸，以增加結腸鏡手術的品質。實驗結果顯示，我們提出的方法能達到盲腸到達率判讀87.98%的準確率，以及能達到90.66%的靈敏度以及86.60%的特異度。
隨著高效能數位系統單晶片需求的增長，數位延遲鎖相迴路和靜態隨機存取記憶體是必不可缺少的電路，但在摩爾定律的影響下，製程電壓溫度的變異將會大幅影響電路的效能，本篇論文提出追蹤電路設計，能夠讓延遲鎖相迴路和靜態隨機存取記憶體擁有抗製程電壓溫度變異。在延遲鎖相迴路中提出的相位追蹤產生電路僅在2個週期內就產生兩個追蹤上升和下降相位，以實現快速鎖定，並且還能操作在寬頻上，此寬頻的全數位延遲鎖相迴路的操作頻率為160MHz至2GHz，測得的峰值抖動為6.89ps和16.67ps，該晶片採用台積電90nm CMOS製程製造。在靜態隨機存取記憶體中，其追蹤電路能實現具有自適應字元線控制技術，可自動調節讀寫字元線的脈衝寬度，以實現抗製程電壓溫度變異並能降低切換功率。該晶片採用台積電90nm CMOS製程製造，其1-Kb 7T SRAM在0.4 V的操作頻率為11.6 MHz時，其量測結果之最低平均功率消耗為4.76 pJ。該電路設計於0.4至1 V達到穩定且低功耗，並可廣泛用於超低電壓架構。

關鍵字

機器學習 ； 深度學習 ； 重度憂鬱症 ； 腦電圖 ； 重複性經顱磁刺激 ； 間歇性脈衝式經顱磁刺激 ； 盲腸到達率 ； 大腸鏡 ； 影像辨識 ； 追蹤電路 ； 全數位延遲鎖相迴路 ； 快速鎖定 ； 靜態隨機存取記憶體 ； 超低電壓 ； 抗製程電壓溫度變異

並列摘要

With the continuous expansion of big data, cloud computing and high-performance computers, the breakthrough of artificial intelligence technology has set off a third wave of development, and the development potential of artificial intelligence technology in the field of healthcare is very huge. In this thesis, artificial intelligence is applied to recognize Electroencephalography (EEG) and colonoscopy image. Major depressive disorder (MDD) is increasingly recognized as a chronic, deteriorating illness with high comorbidity. A significant proportion of patients with MDD fail to respond to sequential antidepressants. Such treatment-resistant depression can be treated with noninvasive brain stimulation, such as repetitive transcranial magnetic stimulation (rTMS) and intermittent theta-burst stimulation (iTBS). We analyzed EEG signals from patients with MDD. Antidepressant responses were predicted by observing the features of patients with MDD receiving rTMS or iTBS treatments. Machine learning is proposed to distinguish responders of rTMS and iTBS. The verification accuracy rates are 92% and 90.9%, respectively. Colonoscopy is the best way to prevent the colorectal cancer (CRC) nowadays. However, it is a highly operator-dependent examination. Therefore, quality assurance and surveillance are quite important. Poor cecal intubation rate is correlated to increase risk of post-colonoscopy CRC. In order to improve the quality of medical care, this thesis propose deep learning for recognizing cecal intubation. This proposed method made sure success of cecal intubation for each examination, thereby increasing the quality of colonoscopy. Experimental results show that our proposed method obtains 87.98% accuracy rate, and achieves a sensitivity of 90.66% and a specificity of 86.60%.
With the growing requirement of high performance digital system on chip (SOC), all digital delay-locked loops (DLL) and static random-access memory (SRAM) are indispensable circuits. Under the influence of Moore's Law, variations in process voltage and temperature will greatly affect the circuit's performance. We proposed a tracking circuit design that can make the DLL and SRAM resistant to process-voltage- temperature (PVT) variations. The proposed phase tracking generator (PTG) of all digital DLL (ADDLL) produces two tracking rising and falling phases in only 2 cycles for fast lock and wide-range. The wide-range ADDLL operates from 160MHz to 2GHz. The measured peak-to-peak jitters are 6.89ps and 16.67ps at 2GHz and 160MHz. This chip is fabricated in TSMC 90nm CMOS technology. In SRAM, the tracking circuit is designed to realize the adaptive wordline control technology, which can adjust the write and read wordline pulse width automatically to achieve PVT variation tolerance and to reduce switching power consumption. The SRAM is fabricated in the TSMC 90-nm CMOS process. The measurement results indicate that minimum average energy consumption per access of 4.76 pJ can be adopted with operation frequency of 11.6 MHz at 0.4 V. This proposed design achieves stable and low-energy consumption from 0.4 to 1 V, which can be widely adopted in energy-constraint applications.

並列關鍵字

Machine Learning ； Deep Learning ； Major Depressive Disorder (MDD) ； Electroencephalography (EEG) ； Repetitive Transcranial Magnetic Stimulation (rTMS) ； Intermittent Theta-Burst Stimulation (iTBS) ； Cecal Incubation Rate (CIR) ； Colonoscopy ； Image Recognition ； Tracking Circuit ； All Digital Delay-Lock Loop (ADDLL) ； Fast Locking ； Static Random Access Memory (SRAM) ； Ultra-Low Power ； Power-Voltage-Temperature (PVT) Variation Tolerance

參考文獻

[1] W. Knight, “The Dark Secret at the Heart of AI”, MIT Technology Review.

[2] D. Sussillo et al., “Opening the Black Box: Low-dimensional Dynamics in High-dimensional Recurrent Neural Networks”, Neural Computation, 25, 626–649, 2013.

[3] Frost Sullivan, “Cognitive Computing and Artificial Intelligence Systems in Healthcare,” 2015.

[4] IEEE SPECTRM, “How IBM Watson Overpromised and Under-delivered on AI Health Care,” 2019.

[5] R. Poplin et al., “Prediction of Cardiovascular Risk Factors from Retinal Fundus Photographs via Deep Learning,” Nature Biomedical Engineering, vol. 2, no. 3, pp. 158–164, 2018.