27 June 2024
A few corrupted slices can undermine an otherwise useful OCT scan. If an AI model gives those slices too much influence, its prediction for the whole volume can become unreliable. In our Medical Image Analysis paper, we introduced Uncertainty-Based Instance eXclusion (UBIX) to address this problem.
UBIX works with multiple-instance learning: individual slices contribute to a prediction for an entire volume. Before those contributions are pooled, UBIX uses uncertainty estimates to identify suspicious instances and reduce or exclude their influence.
The adjustment happens at inference time. The model does not need to be retrained on the new artifacts. Although we studied OCT slices, the same idea can be applied to other definitions of instances and bags.
We evaluated UBIX for staging age-related macular degeneration. Models were trained using scans from one scanner and tested on external datasets from other vendors, including vendor-specific artifacts.
With UBIX, quadratic weighted kappa decreased from 0.861 to 0.708 on the external data containing artifacts. A state-of-the-art 3D network without UBIX decreased from 0.852 to 0.084 on the same test set. These are results for the evaluated models and datasets: they show how limiting the contribution of unfamiliar artifacts can help preserve performance across scanners.
The source code and trained model weights are available on GitHub. The paper was published online on 27 June 2024 and appeared in the October 2024 issue of Medical Image Analysis. UBIX also features in my PhD thesis.
Please use this when referring to the post or publication:
C. de Vente, B. van Ginneken, C.B. Hoyng, C.C.W. Klaver and C.I. Sánchez. "Uncertainty-aware multiple-instance learning for reliable classification: Application to optical coherence tomography", Medical Image Analysis, 2024;97:103259.