Neural-network-based multimode fiber imaging and position sensing under thermal perturbations

Multimode fibers (MMFs) have a very large number of propagating modes per unit area and therefore allow for imaging with a very large number of pixels relative to their diameter. This makes MMFs perfect candidates for ultrathin endoscopes in applications such as deep brain imaging. However, the accuracy of the input-output relation that is needed, e.g., for distal spot scanning without moving parts, requires a new calibration after the fiber position or temperature has been significantly altered. While neural networks have been used before to attempt to solve these challenges, we present an MMF-based imaging method that tolerates and classifies different fiber positions, using two single-layer fully-connected neural networks that only require the optical intensity without measuring the optical phase. One network learns the nonlinear relation between the input and output intensities and allows for image reconstruction in the presence of position changes, while the other network classifies that position change for different images. We show that our method is superior to memory-effect-based position sensing, both for small position changes where the relation between position change and output specklegram rotation angle is linear, as well as for larger position changes where this linearity and uniqueness break down. We also show that the position classification results are robust to temperature and polarization perturbations, and that our position classifier is able to effectively generalize. Likewise, we show that our imaging network also is robust to 30°C perturbations in temperature and 10° in polarization.