Source: Zhongshan Ophthalmic Center
Written by: Zhongshan Ophthalmic Center
Edited by: Tan Rongyu, Wang Dongmei
The team of Prof. Yizhi Liu and Prof. Haotian Lin from Zhongshan Ophthalmic Center, Sun Yat-sen University and Prof. Xiyang Liu from School of Computer Science and Technology, Xidian University, China has taken five years working on creating a novel annotation technique called "Visionome". The research has recently been published in
Nature Biomedical Engineering (IF=17.135), titled "Dense anatomical annotation of slit-lamp images improves the performance of deep learning for the diagnosis of ophthalmic disorders". This technique intelligently and efficiently improves the diagnosis of ophthalmic disorders, and has been put into clinical application.
Previous medical datasets for machine learning were often collected for a single task, such as image-level classification on a specific disease, and therefore led to inadequate data for data mining and meaningful features extractions, reflecting the major bottleneck of the medical annotation for AI training. Moreover, data from most rare diseases is less readily available, undermining the representativeness of medical data, and hindering the development of algorithms. Therefore, the team launched a Medical Artificial Intelligence “Lego” Project, hoping to break through the data heterogeneity barriers of different disease disciplines by converting multidisciplinary medical data into “Lego” modules that can be combined together.
As the first achievement of the Medical Artificial Intelligence ‘Lego’ Project, Visionome has implemented the interdisciplinary and multi-pathological application of artificial intelligence. Inspired by genome sequencing, the team combined genomics with computer vision, and developed “Visionome” to establish a densely annotated dataset, based on anatomical and pathological segmentations. A professional data-annotation team of 25 clinicians using 14 labels described the segmented ocular structures of lesion location, and six pathological lesions based on 54 classification labels were used to describe the pathological features of the segmented ocular lesions. They finally generated 1,772 general classification labels, 13,404 segmented anatomical structures, and 8,329 pathological features.
The workflow of Visionome
"Visionome yielded 12 times more labels than the image-level classification for a single task. It improves the performance of deep learning for the diagnosis of ophthalmic disorders” Prof. Lin stresses that the highlight of Visionome is that it has infinite possibilities to become an excellent “doctor.” Using Visionome, the team created an ophthalmic diagnostic system, the DSV. A user can obtain a comprehensive multi-region diagnostic report by uploading an image of the anterior segment to the DSV within seconds, promoting active healthcare and a shift in the mindset of clinicians and patients who entrust clinical care to machines.
DSV clinical application
In the next step, the team aims to utilize blockchain technology in healthcare on a large scale to advance the Medical Artificial Intelligence "Lego" Project across more diseases. They believe that the advantages offered by blockchain address the shortcomings of traditional data storage, namely the rigorous requirements for data security that currently hinder information sharing, as well as data ownership verification. Utilizing blockchain technology in combination with Visionome has promising prospects for the future of the healthcare industry.
Link to the paper:
https://www.nature.com/articles/s41551-020-0577-y
