On hallucinations in tomographic image reconstruction
This work addresses the critical issue of false structures (hallucinations) in medical imaging for practitioners using deep learning-based reconstruction methods, which could lead to misdiagnoses.
This paper investigates the phenomenon of hallucinations in tomographic image reconstruction, which arise from inaccurate prior information learned by deep neural networks. It proposes decomposing image estimates into generalized measurement and null components and introduces a hallucination map to understand the prior's effect in regularized reconstruction methods.
Tomographic image reconstruction is generally an ill-posed linear inverse problem. Such ill-posed inverse problems are typically regularized using prior knowledge of the sought-after object property. Recently, deep neural networks have been actively investigated for regularizing image reconstruction problems by learning a prior for the object properties from training images. However, an analysis of the prior information learned by these deep networks and their ability to generalize to data that may lie outside the training distribution is still being explored. An inaccurate prior might lead to false structures being hallucinated in the reconstructed image and that is a cause for serious concern in medical imaging. In this work, we propose to illustrate the effect of the prior imposed by a reconstruction method by decomposing the image estimate into generalized measurement and null components. The concept of a hallucination map is introduced for the general purpose of understanding the effect of the prior in regularized reconstruction methods. Numerical studies are conducted corresponding to a stylized tomographic imaging modality. The behavior of different reconstruction methods under the proposed formalism is discussed with the help of the numerical studies.