In this episode, I sit down with Dr. Ibrahim Ethem Hamamci to unpack what it really takes to build modern AI for 3D CT—covering multimodal learning (images + text), the practical realities of clinical deployment, and why “moving beyond 2D” is such a big shift in radiology AI.

We dive into CT-RATE, a large-scale dataset of chest CT volumes paired with radiology reports (plus labels/metadata), why it’s been such a catalyst for 3D vision–language research, and what it enables for training foundation-style models in radiology. The dataset is publicly available on Hugging Face here: https://huggingface.co/datasets/ibrahimhamamci/CT-RATE

If you want a clear primer on core concepts in AI for radiology—representation learning, multimodal supervision and what changes when you go from 2D images to 3D volumes—this conversation is a great starting point. We also dig into why a lot of work still needs to be done: models need to be more robust across scanners and sites, evaluation has to go beyond headline metrics, and real clinical use demands reliability, workflow fit, and accountability.

In this episode, I speak with Mohammed Baharoon, MMSc candidate in Biomedical Informatics at Harvard Medical School and researcher in the Rajpurkar Lab. We discuss his latest work at the intersection of radiology AI, dataset creation, and AI-powered medical education.

Our conversation focuses on two of his recent projects: ReXGroundingCT, an open-source dataset that links free-text CT reports with pixel-level segmentations to improve grounding and explainability in radiology AI and RadGame, a new interactive, gamified platform designed to enhance radiology education using AI-driven case generation and feedback. We explore how these ideas developed, what challenges he encountered while building them, and how they fit into the broader future of trustworthy and effective medical imaging AI.

ReXGroundingCT: https://arxiv.org/abs/2507.22030

RadGame: https://arxiv.org/abs/2509.13270

radgame.io

In this first episode of Pixels 2 Patients, I’m joined by Max Ilse, Senior Researcher at Microsoft Health Futures, to discuss his latest paper, Data Scaling Laws for Radiology Foundation Models (arXiv:2509.12818).

We dive into the core findings of his research, exploring how scaling laws, long studied in natural language processing and vision, apply to the medical imaging domain. We talk about the trade-offs between building large, generalist models that perform across many tasks versus developing highly specialised models that excel within a particular population or modality.

Max also shares his thoughts on the current bottlenecks facing radiology foundation models, underexplored areas of research in the field, and what he sees as the next frontiers in scaling and deploying AI systems for healthcare.

If you’re interested in where medical imaging AI is heading — and what it takes to move from scaling experiments to clinical impact — this is a conversation you won’t want to miss.