- Joined
- Apr 2, 2007
- Messages
- 1,060
- Reaction score
- 118
I figured I'd give a little insight into my experience with what's happening in AI/ML and radiology. There's a great opinion piece from the New York Times that came out today. It echoes my thoughts and prompted me to write this post.
The gestalt of the article is that deep learning has enabled progress on several once-thought-to-be-impossible challenges, but with the same essential limitation as machine learning models of old.
Let me give a real-world example from speech recognition.
Speech recognition plateaued in performance between 1990 and 2010. Despite computer speeds doubling along Moore's Law, speech recognition models would still make the same, infuriating mistakes regardless of processing power. If you've ever used PowerScribe, you know what I'm talking about. On a daily basis, "Low lung volumes" is misinterpreted as "The lung volumes." PowerScribe's flagship product was/is the best in the medical dictation industry, but has long relied on older machine learning models.
Nuance is, of course, updating their system to use deep learning. Deep learning models broke plateaus in nearly every field, including speech recognition. Try going to Google.com and clicking on the microphone. Google's voice recognition is significantly better that Nuance's, even for many complex medical terms!
The problem is, both Nuance and Google's systems are still just for voice transcription. They don't make pizza, or fly helicopters, or distinguish images of cats from dogs. The models do one thing - voice to text, and that's it. Those are obviously contrived examples, but what if you wanted the model to do something tangentially related to transcription? Like, transcribing Spanish? Or, translating English to Spanish? Or, just recognizing a word like "ginormous" that it hasn't been trained on? Models can't make inferences outside their learning task.
Herein lies the rub. Each deep learning model is tuned to a specific task. Many achieve super-human performance on that given task. This is not without utility or clinical significance; a pneumothorax detector that could reliably find basilar pneumos on semi-upright films would be very useful, and likely save a few chest tubes from being removed prematurely.
I could give more examples, but I started writing this rather late at night. My thoughts on AI are more conservative than in the past. I think we will see some tremendous algorithms helping radiologists with fairly arduous or time-consuming tasks, but the problem of generalizability is daunting. Deep learning is rooted in calculus, linear algebra and probability theory, and these fields belie "intelligence." Neural networks trained using GPUs are simply more advanced model that are ultimately limited by the structure and parameters of their network and the generalizability of the training data.
The gestalt of the article is that deep learning has enabled progress on several once-thought-to-be-impossible challenges, but with the same essential limitation as machine learning models of old.
Let me give a real-world example from speech recognition.
Speech recognition plateaued in performance between 1990 and 2010. Despite computer speeds doubling along Moore's Law, speech recognition models would still make the same, infuriating mistakes regardless of processing power. If you've ever used PowerScribe, you know what I'm talking about. On a daily basis, "Low lung volumes" is misinterpreted as "The lung volumes." PowerScribe's flagship product was/is the best in the medical dictation industry, but has long relied on older machine learning models.
Nuance is, of course, updating their system to use deep learning. Deep learning models broke plateaus in nearly every field, including speech recognition. Try going to Google.com and clicking on the microphone. Google's voice recognition is significantly better that Nuance's, even for many complex medical terms!
The problem is, both Nuance and Google's systems are still just for voice transcription. They don't make pizza, or fly helicopters, or distinguish images of cats from dogs. The models do one thing - voice to text, and that's it. Those are obviously contrived examples, but what if you wanted the model to do something tangentially related to transcription? Like, transcribing Spanish? Or, translating English to Spanish? Or, just recognizing a word like "ginormous" that it hasn't been trained on? Models can't make inferences outside their learning task.
Herein lies the rub. Each deep learning model is tuned to a specific task. Many achieve super-human performance on that given task. This is not without utility or clinical significance; a pneumothorax detector that could reliably find basilar pneumos on semi-upright films would be very useful, and likely save a few chest tubes from being removed prematurely.
I could give more examples, but I started writing this rather late at night. My thoughts on AI are more conservative than in the past. I think we will see some tremendous algorithms helping radiologists with fairly arduous or time-consuming tasks, but the problem of generalizability is daunting. Deep learning is rooted in calculus, linear algebra and probability theory, and these fields belie "intelligence." Neural networks trained using GPUs are simply more advanced model that are ultimately limited by the structure and parameters of their network and the generalizability of the training data.
