What does it take for a machine to meaningfully represent a biological system?

“How can we build an agentic AI system that recommends biopsy when clinical evidence is incomplete or conflicting?”

We translate a clinical guideline into an executable policy graph. Each node poses a focused clinical question. Specialist agents evaluate the available evidence to answer these questions. An orchestrator traverses the graph and coordinates the agents to reach a decision. Every step produces an auditable trace. This trace shows which variables influenced the decision and how confident the system was.

“Can biological knowledge guide pretraining for metastasis prediction?”

Predicting axillary lymph node metastasis is difficult because labelled datasets are small. Prior research suggests that the tissue surrounding a tumour or lymph node contains signals associated with metastasis. We use this biological insight to design a pretraining objective for breast ultrasound. A ViT-B model learns to predict the Sobel edge response within masked peritumoral regions. The resulting encoder outperforms state-of-the-art methods for predicting metastasis on small datasets.

“How can we compress an ultrasound foundation model without sacrificing diagnostic accuracy?”

We distil knowledge from ultrasound foundation models into small ViT models using a JEPA-style latent prediction objective. During distillation, we observe a collapse in the effective rank of the learned representations. This collapse limits the diversity of information retained by the student model. We address it by encouraging the model to preserve orthogonal information. The resulting models match or outperform their teachers on several downstream tasks.

“Can a transformer segment tumours without a decoder?”

Most segmentation models pair an encoder with a dedicated decoder. We replace the decoder with a single linear head applied directly to the ViT’s patch tokens. The head predicts the fraction of tumour area within each patch instead of assigning a binary label. With this design, a tiny ViT can localise tumours while achieving performance comparable to a ViT-B model with a decoder.

“Can we build fair models without erasing meaningful biological differences between populations?”

Many domain generalisation methods remove signals that predict cohort membership. However, some differences between cohorts may reflect meaningful biological variation rather than acquisition artefacts. Our radiomics analysis found that most cohort-level differences were concentrated in the wavelet domain rather than the raw image. WAAU-Net acts on this finding by decomposing each image into frequency subbands. It learns how strongly each subband should be normalised. A population discriminator encourages generalisation across cohorts while the selective normalisation preserves useful information.