AI Algorithm Optimization Platform for Medical Image Analytics (ART/303CP)

AI Algorithm Optimization Platform for Medical Image Analytics (ART/303CP)

AI Algorithm Optimization Platform for Medical Image Analytics (ART/303CP)
31 / 03 / 2020 - 30 / 09 / 2021


1. A reference design of the AI algorithm optimization platform for medical image analysis a) Development of generative adversarial network (GAN), active learning algorithms for data generation and annotation collection b) Development of weak supervised learning, explainable ML based solutions for high accuracy DL model building c) Applying the sparse optimization technology to deploy DL model on edge computing device 2. A reference design of building optimized Deep Learning based AI functions for medical image analytics a) AI functions design of object detection and identification for lesion screening b) AI functions design of semantics segmentation and quantitative analysis for disease severity rating 3. A Deep Learning based Computer-Aided Diagnosis application solution a) AI empowered medical device solution for cancer screening Customer Deliverable: a) The initial version of AI-Engine system design for white light gastrointestinal (GI) endoscopy EGC screening. b) The enhanced version of AI-Engine system design for white light GI endoscopy EGC screening. c) The integrated AI-Engine system for EGC screening and Image enhancement algorithm. Algorithms performance indicator for EGC screening: 1. The abnormal region detection under white light endoscopic image: sensitivity >= 80% specificity >= 90% 2. The EGC identification under white light endoscopic image: sensitivity >= 70% specificity >= 70% The average diagnostic level of human experts under white light endoscopic image given in the medical research papers listed in II.1.7[performance benching] has been used as the performance benchmark. The processing speed of algorithms: 1. The abnormal region detection: >=24fps@1080p 2. The EGC identification algorithm: >=24fps@1080p Under the following machine configuration definition: > CPU: Intel(R) Core (TM) i9-9900k CPU (8 Cores, 16 Threads @3.6GHz) > GPU: RTX 2080Ti*2 or stronger computing power > PSU: >= 1200W > RAM: >= 32GB > Storage: 1TB SSD + WD Gold 2TB * 2 > Mother board: ASUS ROG MAXIMUS XI Gene

GT Medical Systems Limited [Sponsor]
Richen Sight (HK) Co. Limited (Contract Service) [Sponsor]

Artificial intelligence (AI) assisted medical image analysis and diagnosis has become more and more popular. AI can learn features, get insight from large quantity of data, do prediction and support repetitive jobs. It can effectively solve the problems of low efficiency of diagnosis, insufficient doctors and inevitable misdiagnosis. At present, in the field of radiology and retina fundus imaging, AI-aided diagnosis has achieved prominent result. In the field of digestive endoscopy and digital pathology, with the growth of the market, the demand for AI is becoming more and more urgent. However, due to the high cost of data acquisition, the weak annotation data and the real-time requirement of clinical application, the real landing of AI medical image analysis still faces great challenges.

This proposal aims at enhancing the development of Computer-Aided Diagnosis (CAD) in the medical domain. We will build an AI algorithm optimization platform to mitigate the impact of insufficient data, high annotation cost and inaccurate annotation on the model accuracy and speed up the processing through sparse optimization technology. We will focus on developing Deep Learning (DL) based CAD solutions for digestive tract endoscope image analysis. Leverage on the optimization framework to provide model design with higher sensitivity and specificity, and achieve real-time processing with limited computing resources.