Scientists Pull Their Own Medical AI Study, Citing Doubts
A reminder that flashy AI health claims can collapse — even their authors agree.
Researchers proposed a model called SegKAN for segmenting hepatic vessels in CT scans, where vessels often suffer from image fragmentation and noise interference. The model uses a novel convolutional network structure for image embedding, aimed at smoothing image noise and preventing gradient explosion, and transforms spatial relationships between Patch blocks into temporal relationships to address capturing positional relationships in traditional Vision Transformer models. On a hepatic vessel dataset, the Dice score improved by 1.78% compared to the existing state-of-the-art model. The paper was later withdrawn; the first author's comment states that due to unresolved limitations in the study, the conclusions are not yet sufficiently supported, and advises readers not to cite it, confirmed by the corresponding author. No PDF is available, and there is no license for the withdrawn version.
- Researchers claimed their AI read CT scans more accurately than existing tools, then withdrew the study saying the results weren't solid enough.
- The improvement was tiny anyway — about 1.78% on one accuracy measure — but such gains often get hyped as breakthroughs.
- Withdrawal is a sign the system worked, but these retractions usually happen quietly, so patients rarely hear about them.
Why It Matters
Medical AI tools are entering hospitals now. Withdrawn studies mean you should ask what's actually proven.