@dr-alwin-chuan.bsky.social
2nd-gen AI could use different methods; eg. dynamic segmentation models. Extreme frequency (20-45MHz) US is just around the corner - we can start including these high res images into our dataset too. Perhaps this will deliver real clinical value. It’s an exciting time to be working in this space
April 4, 2025 at 1:29 AM
2nd-gen AI could use different methods; eg. dynamic segmentation models. Extreme frequency (20-45MHz) US is just around the corner - we can start including these high res images into our dataset too. Perhaps this will deliver real clinical value. It’s an exciting time to be working in this space
Nerve US relies on dynamic scanning. Anatomical variability and low resolution also leads to poor ground truth even amongst expert anaesthesiologists. Finally, our training dataset is limited to 100s of images, compared to 100,000s used in other AI.
2/3
2/3
April 4, 2025 at 1:29 AM
Nerve US relies on dynamic scanning. Anatomical variability and low resolution also leads to poor ground truth even amongst expert anaesthesiologists. Finally, our training dataset is limited to 100s of images, compared to 100,000s used in other AI.
2/3
2/3
At the ASURA conf, I spoke that AI allows novices to confidently diagnose obstetrics, lung, gastric & cardiac ultrasound (US). But for nerve blocks: AI performs well on easy but not in hard anatomy
Why? 1st-gen AI used convolutional networks, U-Net & model tuning, but perhaps we shouldn’t
1/3
Why? 1st-gen AI used convolutional networks, U-Net & model tuning, but perhaps we shouldn’t
1/3
April 4, 2025 at 1:29 AM
At the ASURA conf, I spoke that AI allows novices to confidently diagnose obstetrics, lung, gastric & cardiac ultrasound (US). But for nerve blocks: AI performs well on easy but not in hard anatomy
Why? 1st-gen AI used convolutional networks, U-Net & model tuning, but perhaps we shouldn’t
1/3
Why? 1st-gen AI used convolutional networks, U-Net & model tuning, but perhaps we shouldn’t
1/3