AI Translations in Radiology: A Missed Diagnosis?

If you're banking on AI to bridge language gaps in radiology, think again. A recent study comparing Japanese translations of chest CT reports reveals a significant divide in opinion between AI models and human radiologists.

AI vs. Human: The Translation Battle

In an analysis of 150 chest CT reports, LLM-generated translations were put to the test against human-edited versions. The study enlisted DeepSeek-V3.2, which, along with Mistral Large 3 and GPT-5, acted as an AI judge. They evaluated translations based on terminology accuracy, readability, overall quality, and how 'radiologist-like' they were.

Here's where it gets interesting. The AI judges overwhelmingly favored their own output, with preferences running between 70% and 99% across all criteria. They deemed their translations more authentic in over 93% of cases. But the radiologists? They barely agreed with each other, let alone the AI.

Radiologists Weigh In

Radiologist 1 saw terminology accuracy as largely equal but leaned towards AI for readability and quality in just over half the cases. Radiologist 2, however, found readability mostly on par but leaned towards human edits for quality. Their agreement was almost non-existent, with a QWK (Quadratic Weighted Kappa) of 0.01 to 0.06 between them.

So, what gives? Why the disconnect? It's clear the AI models are confident in their capabilities, but radiologists aren't buying it. Is it because AI lacks the nuanced understanding that a trained human eye provides?

The Reality Check

Automated translation might seem like the future, but can we really trust it without human oversight? The study suggests otherwise. LLM-generated translations might be natural and fluent, but educational accuracy in radiology, AI can't go it alone.

So, where does this leave us? It leaves us with a glaring reality check. You can't replace expert human judgment with algorithms, at least not yet. If you haven't realized the importance of keeping seasoned radiologists in the loop, you're missing the forest for the trees.

AI translations are a leap forward, but they're not ready to run solo in critical fields like medicine. The technology's promising, but until it matures, expert review remains indispensable. Solana doesn't wait for permission, but in radiology, maybe it should.