Autodidactic dense anatomical models

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Not reported in abstract — qualitative claims of 'advanced capability' only

• Primary Outcome: Zero-shot anatomy understanding via Eve-v2 embeddings, demonstrating localizability of anatomical structures and encoding of part-whole relations without task-specific supervision

• Nnt Or Sensitivity: Not applicable — no diagnostic accuracy metrics, sensitivity, specificity, or segmentation Dice scores reported in the abstract; downstream task performance metrics require full paper review

• Confidence Interval: Not reported

Clinical Application

Feasibility in clinical deployment is currently theoretical. The framework requires substantial computational infrastructure for pre-training. Downstream fine-tuning for specific clinical tasks (e.g., organ segmentation, lesion detection) would require integration into existing PACS/RIS workflows. Few-shot learning capability, if validated, could reduce annotation burden for rare conditions — a meaningful practical advantage. Australia's radiology workforce faces significant demand pressures, and AI-assisted anatomical analysis tools have potential utility in supporting reporting efficiency and triage. The TGA regulates AI/software as medical devices under the Software as a Medical Device (SaMD) framework — any clinical deployment of Adam-v2-derived tools would require TGA conformity assessment. The RACGP and Royal Australian and New Zealand College of Radiologists (RANZCR) have published position statements on AI in medical imaging emphasising the need for prospective clinical validation before adoption. PBS implications are not directly relevant at this stage. Australian datasets (e.g., from public hospital networks) are not represented in the training data as described, raising questions about performance on Australian patient demographics and imaging protocols. Radiologists, clinical imaging specialists, and medical AI developers working with CT, MRI, or other cross-sectional imaging modalities requiring anatomical structure recognition, segmentation, or anomaly detection

Abstract

Humans effortlessly interpret images by parsing them into part-whole hierarchies. Yet, deep learning models, despite excelling at capturing multi-level features, often fail to explicitly encode these part-whole hierarchies-an essential aspect of medical imaging, which boasts anatomical hierarchies in nature. To address this limitation, we introduce Adam-v2, a self-supervised learning framework that explicitly learns to encode inherent part-whole hierarchies within medical images through three key branches: (1) "localizability", which acquires discriminative representations to distinguish different anatomical structures; (2) "composability", which learns each anatomical structure in a parts-to-whole manner; and (3) "decomposability", which comprehends each anatomical structure in a whole-to-parts manner. Our extensive experiments showcase Adam-v2's advanced capability in anatomy understanding, unveiled through its embeddings (Eve-v2). Particularly, Eve-v2 demonstrates a zero-shot understanding of anatomy as revealed through its learned properties: ➀ preserving localizability of anatomical structures and ➁ encoding part-whole relations of anatomical structures as well as its emergent properties: ➂ understanding anatomical layouts via interpolation and extrapolation, ➃ associating each image pixel with (dense) semantic embeddings, ➄ recognizing anatomical symmetries, ➅ generating embeddings consistent across scales, and ➆ matching anatomical structures across images of the same patient with different diseases, images of different patients, and augmented views of the same image. Adam-v2 also offers robust and generalizable representations and stands out in ➇ few-shot learning, ➈ full-transfer learning, and ➉ novelty and anomaly detection. These capabilities and performance directly stem from our crafted anatomy learning strategy, which explicitly constructs hierarchies for distinct anatomical structures from unlabeled medical images. Project page: GitHub.com/JLiangLab/Eden.

References

1. 1.Hosseinzadeh Taher, M. R., Gotway, M. B., & Liang, J. (2026). Autodidactic dense anatomical models. Medical Image Analysis. https://doi.org/10.1016/j.media.2026.104117