Page 122 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

2 Use Case Description

2�1 Description

ChatZOC Ophthalmology Large Model is a specialized medical AI model designed by
Zhongshan Ophthalmic Center (ZOC), Sun Yat-sen University. It is to enhance the intelligence
and efficiency of ophthalmic diagnosis, patient services, scientific research, medical education,
and hospital management. By leveraging Natural Language Processing (NLP) and deep learning
technologies in image analysis, ChatZOC offers intelligent consultations, precise screening,
diagnostic assistance, and research support.

ChatZOC pioneered dense anatomical annotation for ophthalmic medical images, significantly
improving data utilization. A self-developed Visionome model can recognize 22 anterior
segment pathologies and provide diagnostic recommendations by integrating visual and
textual data. On the language side, ChatZOC employs multi-level knowledge retrieval and
patient intent recognition to deeply understand complex clinical questions and perform
medical reasoning. It was also implemented an intelligent question rewriting mechanism to
enhance input precision and improve response interpretability.

It effectively addresses critical challenges such as the shortage of ophthalmologists and
healthcare disparities between urban and rural areas. Compared to traditional, labor-intensive
diagnostic approaches, ChatZOC enables rapid responses, standardized diagnostic protocols,
and large-scale screening capabilities. However, there remains room for improvement in
recognizing complex conditions, providing personalized treatments plans, and managing
high computational demands. Moving forward, ChatZOC will continue to evolve and innovate,
contributing to the sustainable development of global eye health.

Supplemental Information:

In response to the questions raised by the judges of ITU, we provide the following information

1) how to include expert supervision and audit?

Response: Firstly, regarding the data aspect, we utilized professional data that was rigorously
standardized by professional doctors. In terms of the model, we established a RAG (Retrieval
Augmented Generation) database and have corresponding supervision mechanisms for the
output. Expert supervision is integral to the ChatZOC framework. ChatZOC serves as an assistive
tool designed to support clinicians, not replace their judgment. Ophthalmologists retain final
decision-making authority and perform audits on AI-generated diagnostic reports to ensure
clinical safety, accuracy, and appropriate application.

2) how is the preparation and labelling of data done in this framework?

Response: The data used for training is medical data from real doctor-patient interaction
scenario. Before commencing annotation, a clear and specific annotation guideline needs
to be formulated to ensure a common understanding of the standards among all personnel
involved. To minimize the influence of personal bias, a double-blind annotation method can be
employed, where each data point is independently annotated by two junior doctors, followed
by review and arbitration by a senior doctor with 20 or more years of experience, ensuring
annotation consistency and accuracy. Upon completion of annotation, the data also undergoes
validation and correction to ensure its completeness and accuracy.

86