Page 117 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

classification and grading within these subspecialties. This deployment provides valuable real-
world validation and data for further refinement.

Use of Foundational Models: The RuiPath model utilizes a foundational model. Specifically,
DeepSeek R1 was employed, primarily through model distillation techniques during the 4.1-Healthcare
training process. The deployment is based on DCS AI Solution, incorporating optimization
techniques such as KV cache offloading and MoE (Mixture of Experts) expert prefetching. These
optimizations have reportedly reduced the inference cost of the DeepSeek R1 component by
50%.

The specific benefits of the solution are as follows:
1. Breadth: Covers 90% of the cancers in the world every year.
2. Depth: The depth of sub-speciality knowledge Q&A reaches the expert level, and the
Q&A accuracy exceeds 90%, which is the leading level in the world.
3. Efficiency: In the past, it took about 10 minutes to diagnose a tissue slice under a
microscope. Now, after AI interactive reading, AI calculation time for a single slice is only
seconds, and the overall diagnosis efficiency is several times.

Use Case Status: The AI-Enabled Pathology Model is currently being pilot tested in 11
subspecialties at Ruijin Hospital, including breast, prostate and thyroid pathology. It will be
officially implemented and operated at Ruijin Hospital in the latter half of 2025.

Partners: Ruijing Hospital

2�2 Benefit of use case

The DCS AI solution provides the following benefits:

(1) Tool-based data engineering shortens the data preparation period for medical training by
80%.

(2) System-level model training and inference acceleration capabilities shorten the model
training period by 30% and double the inference concurrency.

(3) Simplified application development platform, enabling non-professional developers to
quickly get started. This programme embodies a commitment to industry, innovation and
infrastructure, driving advances in cancer diagnostics and healthcare automation.

It leverages artificial intelligence technologies to deliver intelligent consultation and precise
screening, improving the rate of early diagnosis and treatment efficiency of common diseases
while reducing misdiagnosis rates. It helps cover the shortage of medical resources and raises
global standards for public health.

In the realm of personnel efficiency, the transition from manual specimen information input and
allocation of 2 staff members per lesion tissue to utilizing AI voice input and human verification
has reduced the requirement to just 1 staff member per lesion tissue. Regarding diagnostic
efficiency, initial reviewing physicians now spend 10-15 minutes per patient with AI-assisted
image analysis and report editing, down from 20-40 minutes. Similarly, confirmatory physicians'
time per patient has decreased to 10 minutes with AI prompting for review and automatic
input, a significant improvement from the previous 15-30 minutes spent on manual review
and report modification. These innovations not only boost workforce productivity but also

81