Simplifying Adaptive Optics in Laser-Scanning Microscopy via Deep Learning and Detector Array

Due to its versatility and information-gathering capabilities, confocal laser-scanning microscopy (CLSM) is indispensable for investigating specimen structure and function at both cellular and sub-cellular levels. More recently, CLSM has evolved into what is known as image-scanning microscopy (ISM). In ISM, a fast detector array replaces the traditional single photon detector and pinhole. Each element of the detector array acts as a virtual pinhole, as for confocal microscopy, but without photon rejection. This data acquisition allows for the reconstruction of images with spatial resolution beyond the diffraction limit, as for an ideally infinite small pinhole in CLSM, but with an excellent signal-to-noise ratio. Notably, in ISM, the optical sectioning capability of CLSM is preserved, guaranteeing three-dimensional imaging in thick specimens, such as tissues or whole organisms. However, as with all optical microscopy techniques, the resolution is often adversely affected by the optical properties of the optical elements of the system or the specimen itself. This is particularly relevant in thick samples, where optical aberrations limit imaging depth. Adaptive optics (AO) has emerged as the most popular solution to compensate for optical aberrations. These approaches involve spatially structuring light to pre-compensate for distortion introduced by aberrations. Inspired by advancements in related fields such as telecommunications and astronomy, many important technological advancements in beam-shaping tools have improved adaptive optics. Yet, aberration identification remains a critical issue. The complexity of the hardware approaches or the request of long and sample-invasive software iterative approaches to identify aberration makes current adaptive-optics microscopy implementation not user-friendly. Here, I propose a simple and gentle approach for optical aberration correction based on an image-scanning microscope and conventional beam-shaping elements. By this approach, I demonstrate how the information the detector array provides is used to implement a new class of AO laser scanning microscopy architectures that offers access to aberrations. I built an ISM system with an AO element (Deformable Mirror) and an array detector composed of 5x5 elements. In short, the detector array is able to provide a four-dimension dataset: a two-dimensional microimage i(xd | xs), with xd = (xd, yd) the detector space, for each point of the scanning space xs = (xs, ys). Alternatively, this dataset can be viewed as a collection of scanned images i(xs | xd), one for each array element. This new dataset effectively encodes the optical aberrations induced by the optical elements and the specimen. To decode the optical aberration information from the experimental dataset, I calculated the phase correlation between the scanned images resulting in a series of correlograms. Subsequentially, I developed a custom convolutional neural network (CNN) designed to directly decode the aberration information in terms of Zernike coefficients from these correlograms. Our novel imaging technique, in principle, estimates aberrations from a single acquisition, enabling real-time correction with an AO element. In conclusion, the work presented in this thesis paves the way for a real-time and versatile adaptive optics architecture capable of accurate and efficient aberration estimation directly from experimental images, thereby unlocking new avenues for deep imaging.