什維新智醫(yī)療科技(上海)有限公司2020年5月，攜手英國白金漢大學(xué)在上海發(fā)布超聲影像智能輔助診斷系統(tǒng)。

我們現(xiàn)有解決方案的*性和技術(shù)*性：

•能提供輔助意見的自動癌癥識別系統(tǒng)，幫助及早發(fā)現(xiàn)癌癥，增加成功、及時治療的機(jī)會。

•利用*的深度學(xué)習(xí)神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)，從二維超聲圖像中識別病變狀態(tài)，其準(zhǔn)確度可達(dá)到甚至超過有經(jīng)驗的超聲醫(yī)生；

•在訓(xùn)練深度學(xué)習(xí)神經(jīng)網(wǎng)絡(luò)模型的時候，選取了醫(yī)院常用的各種型號的超聲機(jī)采集的圖像，保證訓(xùn)練模型在臨床實踐中有更好和更廣泛的適用性；

•為不同器官（如甲狀腺、乳腺、肝、卵巢和膀胱）的腫瘤分類開發(fā)一個通用的深度學(xué)習(xí)神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)，有望建立更可靠的基于機(jī)器的診斷模型，并易于技術(shù)維護(hù)和支持；

•采用計算機(jī)視覺領(lǐng)域的全新技術(shù)，使用深度學(xué)習(xí)神經(jīng)網(wǎng)絡(luò)檢測主要腫瘤特征，以協(xié)助醫(yī)生對腫瘤特征的描述，并為醫(yī)生的結(jié)果診斷決策提供支持。

我們未來解決方案的*性和技術(shù)*性：

•開發(fā)一種新的基于人工智能的軟件產(chǎn)品，該產(chǎn)品可以自動分析二維超聲圖像，對不同類型的腫瘤可以自動進(jìn)行腫瘤識別、檢測、分割和診斷決策。

•*設(shè)計和優(yōu)化的深度學(xué)習(xí)神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)，用于從二維超聲圖像中診斷腫瘤狀態(tài)，其準(zhǔn)確度將超過大多數(shù)經(jīng)驗豐富的超聲醫(yī)生，同時大大降低了經(jīng)過培訓(xùn)的深度學(xué)習(xí)神經(jīng)網(wǎng)絡(luò)模型的資源需求，使模型更容易安裝在云甚至移動平臺上；

•基于識別的檢測機(jī)制，具有強(qiáng)化學(xué)習(xí)能力，能夠自動定位和檢測超聲圖像中的感興趣區(qū)域，顯著改善用戶體驗和檢測精度；

•一種自動和/或半自動方法，準(zhǔn)確分割良性和惡性的感興趣區(qū)域（即腫瘤），以便準(zhǔn)確檢測各器官的癌癥跡象；

•強(qiáng)大的腫瘤標(biāo)志物檢測技術(shù)，用于推導(dǎo)和測量腫瘤特性和特征；

•通過對決策流形狀的可視化和拓?fù)浞治?，提高對由深度學(xué)習(xí)神經(jīng)網(wǎng)絡(luò)模型做出的診斷決策的理解能力，并保證診斷決策的可理解性。

與現(xiàn)有的其他圖像模式相比，超聲圖像分析充滿了巨大的技術(shù)挑戰(zhàn)。在現(xiàn)有和未來解決方案的背后，是我們勇于接受挑戰(zhàn)的動機(jī)和決心，我們的愿景是超聲圖像檢查的非侵入性的、安全并且費用低廉的優(yōu)點，能廣泛地應(yīng)用到醫(yī)院、基層醫(yī)療衛(wèi)生機(jī)構(gòu)等，造福大眾。

Uniqueness and technical advances of our existing solutions:

• Developing an Automatic Cancer Recognition System which provides a second opinion to help in early cancer detection and increase the chances for successful and timely treatments. 
• Exploiting advanced deep learning neural network architectures for lesion status recognition from 2D ultrasound images with levels of accuracy that match or outperform experienced radiologists;
• Training deep learning neural network models with images collected from various makes of ultrasound machines commonly deployed in hospitals with promises of better and wider applicability of the trained models in clinical practice;
• Developing a generic deep learning neural network architecture for tumour classification for different types of cancer (such as thyroid, breast, liver, ovary and bladder) with promise of more robust diagnostic machine-based models, and ease of technical maintenance and support;
• Adapting state-of-the-art techniques in computer vision for detecting major tumour signs using deep learning neural network to comprehend, assist doctor’s descriptions of tumour characteristics, and provide support for doctor’s final diagnostic decisions.

Uniqueness and technical advances of our future solutions:

• Develop a novel AI based software product aims to automatically analyse 2D ultrasound image for tumour recognition, detection, segmentation and diagnostic decisions understanding for different types of cancer. 
• Uniquely designed and optimised deep learning neural network architectures for diagnosing tumour status from 2D ultrasound images with levels of accuracy that will outperform most experienced radiologists and specialists, and at the same time greatly reduce resource requirements of the trained deep learning neural network models, making it easier for the models to be installed on cloud or even moving platforms;
• A recognition-based detection mechanism with reinforcement learning capability to automatically locate and detect region of interest in Ultrasound image to significantly improve the user experience as well as detection accuracy;
• An automatic and/or semi-automatic method to accurately segment the region of interest (i.e. tumours) of both benign and malignant natures for accurate detection of cancer signs for various types of cancer;
• Robust tumour sign detection techniques in deriving and measuring tumour properties and characteristics;
• A capability of enhancing the understanding of diagnostic decisions made by deep learning neural network models through visualisation and topological analysis of decision flow shapes with the promise of comprehensibility of the diagnostic decisions.

Behind both the existing and future solutions is our motivation and determination to meet the extreme technical challenges of ultrasound image analysis in comparison to many existing works in other image modalities for utilising the non-invasive, safe and cost effective benefits of ultrasound imagery that is widely available in hospitals and health centres.