1. Reference design of generalized training data preparation for the two specific medical image analytics applications: (a) Development of Deep Learning model with public data for annotation extraction; (b) Development of a user-friendly interface for annotation results review. 2. Reference design of building state-of-the-art Deep Learning model for the two specific medical image analytics applications: (a) Valuation of new annotations generated in 1(a); (b) Development of new Deep Learning models with new large-scale training data; (c) General guideline for Deep Learning model building. 3. Two Deep Learning Computer-Aided Diagnosis application solutions: (a) Digital pathology image analysis using Deep Learning; (b) Anomaly detection in capsule endoscopy video. 4. Contract Service Deliverables for Customer 1: (a) Image quality examination algorithm with source codes for capsule endoscopy video; (b) Gastrointestinal tract division algorithm with source codes for capsule endoscopy video; (c) Spectral image enhancement algorithm with source codes for virtual chromoendoscopy; (d) Redundant image content detection algorithm with source codes for consecutive frames; (e) Gastrointestinal abnormality detection algorithm with source codes for capsule endoscopy video.
Comcsoft (JiangSu) Corporation
Richen Sight (HK) Co. Limited
With an overwhelming medical image data volume generated daily for clinical diagnosis, medical image analysis is in great need of automated and computer-aided processing. The problem is further complicated due to the lack of radiologist/pathologist, rapid increase in workload and rise in imaging complexity. A lot of researches have been done, in both academic and industrial fields, trying to meet the high standard requirements for clinical use. However, existing solutions have not been widely adopted in clinics. On the other hand, Deep Learning generates consistent and promising performance across many domains; this algorithm has drawn great attention in the Machine Learning area. A key success factor of a Deep Learning application is the use of large training data volume, which is often being considered “luxurious” in the medical image analytics domain. This proposal aims at enhancing the development of Computer-Aided Diagnosis (CAD) in the medical domain, and developing methods to tackle the “lack of training data” bottleneck; so as to output state-of-the-art Deep Learning techniques. We proposed a semi-supervised image annotation framework for typical training preparation, and provide a guideline in building Deep Learning models for medical image analytics. We will focus on developing two Deep Learning-supported CAD solutions for digital pathology and video endoscopy respectively; both of which will provide higher sensitivity than current manual diagnosis.