BoneGraph: A Domain-Specialised, Self-Correcting Reasoning System for Bone Science Retrieval, Grounded Inference, and Image Mechanics

Bone science literature encompasses various disciplines such as biology, mechanics, materials science, and clinical medicine. The wealth of this information poses challenges in reliably synthesizing knowledge. While general-purpose large language models (LLMs) can provide fluent responses, they often underrepresent the specific domain of bone science, fail to cite specific evidence, and lack a mechanism for durable correction.

The researchers have developed BoneGraph, a specialized system tailored for bone science. This system is presented as a five-tab web application operating on a shared substrate. The substrate consists of a curated full-text corpus of 7,449 documents embedded into 248,629 passage vectors using the SPECTER2 model, a scientific-paper embedding model, and a bone knowledge graph containing 1,597 concepts and 1,699 causal relations.

The five tabs of BoneGraph serve distinct purposes. The first tab, Chat, offers retrieval-augmented question answering with server-rebuilt inline citations. The second tab, Search, conducts raw semantic retrieval without the involvement of an LLM. The third tab, Reasoning, incorporates a self-correcting loop. This loop employs a deterministic physics check and a literature/knowledge-graph critic to constrain the answer. User feedback becomes a durable, per-user rule, enhancing the system's accuracy over time.

The fourth tab, Vision, features a bone-region classifier trained on frozen BiomedCLIP features. This classifier grounds a vision-language model, ensuring robustness against out-of-distribution inputs. It is further augmented with image-embedding correction memory for enhanced performance. The fifth tab, Mechanics, integrates a previously developed data-driven image mechanics model (D2IM) that predicts displacement and strain fields from a single undeformed micro-CT image.

All inference processes within BoneGraph are performed locally, without relying on third-party API calls. The public beta of the system is accessible at bonegraph.org. The retrieval component of BoneGraph has achieved a mean reciprocal rank (MRR) of 0.928 on a benchmark containing 30 questions across seven domains. The Vision classifier attains an impressive 92.6% accuracy when evaluated on the held-out MURA (MUsculoskeletal RAdiographs) dataset.

A benchmark focused on grounded reasoning demonstrates the effectiveness of BoneGraph. When provided with the correct passage, the system elevates answer accuracy from 42% to an impressive 78%. This significant improvement underscores the system's potential to revolutionize bone-science informatics. To the best of our knowledge, BoneGraph represents the first domain-specialized system to seamlessly integrate curated retrieval, deterministic physics-grounded self-correction, and durable per-user learning specifically tailored for bone science.