Page 118 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

streamline diagnostic processes, aligning with the goals of promoting decent work practices
and fostering economic growth.

DCS AI model project outlines a comprehensive model training process in two phases: first,
compressing and encoding image data, and second, aligning image-text spaces. The model
aims to enhance cancer diagnosis, facilitate similar medical history retrieval, automate pathology
report generation, and drive innovation in education and research. By deploying this pan-cancer
multi-modal pathology grand model, the goal is to achieve five core functionalities, leveraging
cutting-edge technologies like Vision Transformers and multi-modal models. This initiative
embodies a commitment to Industry, Innovation, and Infrastructure, driving advancements in
cancer diagnostics and healthcare automation.

2�3 Future Work

The development roadmap for the RuiPath system includes several key enhancements and
expansions aimed at broadening its capabilities and impact:

• Multimodal Data Integration: Future iterations plan to move beyond the current reliance
on H&E images and pathology report descriptions. The strategy involves incorporating
additional data modalities, such as molecular diagnostics (e.g., genomic sequencing
data) and immunohistochemistry (IHC) results. This integration aims to create a more
comprehensive patient profile, enabling precision diagnostics and treatment planning
across the patient's entire healthcare journey.
• Expanded Downstream Task Coverage: The scope of applications will be extended to
include more complex downstream tasks beyond classification and grading. A primary
focus will be the development and validation of robust prognosis prediction capabilities,
leveraging the integrated multimodal data.
• Second-Generation Model Architecture: A next-generation architecture is under design
to support these advancements. This includes:

1� Enhanced Self-Supervised Learning: Updating the self-supervised algorithms to
effectively process and learn from higher-resolution pathology images.
2� Advanced Multimodal Fusion: Refining the Mixture of Experts (MoE) architecture to
optimally fuse encoded information from diverse data modalities (H&E, molecular,
IHC, etc.), aiming for further improvements in model accuracy and robustness.

• Data Acquisition and Standardization: Recognizing the critical role of data quality and
diversity, future plans include:

1� Deepened Hospital Collaborations: Establishing in-depth partnerships with multiple
healthcare institutions.
2� Standard Development: Collaborating with partner hospitals to jointly define
standardized metrics and scanning protocols for digital pathology slides.
3� Multi-Center Data Collection: Utilizing these standards to collect larger volumes of
high-quality, multi-center data for model training, enhancing generalizability.
• Open-Source Initiative and Expansion: To foster broader adoption and innovation, there
is a plan to open-source relevant aspects of the work in the second half of 2025. This
initiative aims to empower more healthcare providers, particularly primary and community
hospitals, with advanced AI tools for pathology analysis.

82