Uncover This Tech Term: Uncertainty Quantification for Deep Learning
Shahriar Faghani, Cooper U. Gamble, Bradley James Erickson · Korean Journal of Radiology · 2024
What is Uncertainty Quantification?Deep learning (DL) has been recognized for its potential in radiology, yet concerns regarding its reliability in clinical workflows limit its adoption.This has become a greater challenge in radiology societies following the recent awareness of hallucinations in large language models [1].These arise from predictions made without an estimate of the trustworthiness of DL models [2].DL models, including computer vision and language models, generate numerical outputs that resemble probability.However, these outputs are used for training the model and are not indicative of the actual likelihood of a specific outcome, because they lack calibration [3].Consequently, if a model outputs a value of 0.8 for a particular diagnosis, it does not necessarily