- Keywords: surgical data science, multispectral imaging, ambiguity, uncertainty estimation, deep learning, invertible neural networks, out of distribution detection
- TL;DR: We suggest a novel multistage framework for uncertainty handling in multispectral imaging.
- Abstract: Replacing traditional open surgery with minimally-invasive techniques for complicated interventions such as partial tumor resection or anastomosis is one of the most important challenges in modern healthcare. In these and many other procedures, characterization of the tissue remains challenging by means of visual inspection. Conventional laparoscopes are limited by “imitating” the human eye, recording three wide color bands (red, green and blue). In contrast, multispectral cameras remove this arbitrary restriction, and allow for the capture of many narrower bands of light. In previous work, we have shown that the inverse problem of converting pixel-wise multispectral measurements to underlying tissue properties can be addressed with machine learning techniques. Key remaining challenges are related to the handling of uncertainties: (1) Machine learning based approaches can only guarantee their performance on data that is sufficiently similar to the training data, hence, out of distribution (OoD) samples have to be handled and (2) most algorithms provide a point estimate of a physiological parameter (e.g. blood oxygenation sO2), neglecting the fact that the problem may be inherently ambiguous (i.e. two radically different tissue parameter configurations may - in theory - result in similar measured spectra). We tackle these shortcomings with a framework for uncertainty handling that involves (1) filtering OoD samples and then (2) computing the full posterior probability distribution for tissue parameters given the measured spectra. Analysis of the posteriors not only provides us with a means for quantifying the uncertainty related to a specific measurement but also enables a fundamental theoretical analysis about which tissue properties can in principle be recovered with multispectral imaging (MSI).
- Code Of Conduct: I have read and accept the code of conduct.
- Remove If Rejected: Remove submission from public view if paper is rejected.
- Link: https://link.springer.com/article/10.1007/s11548-019-01939-9