Should I have become a radiologist? The hype versus reality of radiology AI (ameyarad.github.io)

A musculoskeletal radiologist recounts a deep-dive into the “hype vs reality” of radiology AI after years of clinical practice and hands‑on experiments: training an X‑ray bone‑age model via transfer learning, building a reporting assistant combining an LLM with automatic speech recognition, and attempting edge‑AI deployments in the NHS. He maps radiology AI onto Gartner’s hype cycle and argues we’re past wild optimism but far from a mature, general solution. While vision‑language models (VLMs) and foundation models promise broad capability, early efforts — including a 2023 open‑source radiology foundation model and a Harvard MRI model trained on ~220k scans (~19M slices) — produced glaring errors and hallucinations, exposing limitations of transformer‑based approaches and massive data/GPU costs. The practical implications matter: most cleared tools are narrow CNNs that assist specific tasks (e.g., mammography) rather than replace radiologists, and cost‑effectiveness is still debated. FDA clearance isn’t clinical approval, conflicts of interest are common in published AI studies, and many hospitals lack the GPUs or cloud options needed for high‑accuracy models. Edge AI requires heavy quantization that reduces performance. Bottom line: radiology AI currently augments workflow for narrow problems but cannot handle the full complexity of clinical imaging. Careful evaluation (eg. CLAIM checklist), robust validation, infrastructure investment, and realistic expectations are essential for safe, useful adoption.