*5.1. Imaging*

Imaging sensors utilize an array of optical sensors such as a CMOS array (complementary metal-oxide-semiconductor array; the most used image sensor for digital cameras). Images of the specimen can be used to identify the target presence and concentration as the molecules exhibit different coloration, fluorescence, or light scattering, with varying morphology and spatial distribution. In this manner, several imaging biosensors have been developed to eliminate the need for labels and bioreceptors.

A growing field of imaging-based biosensors utilizes lens-free imaging techniques [143,144]. Since the images from lens-free imaging are not in focus, computational techniques are needed for image reconstruction, the most common of which is deep learning (mostly based on ANN with "deeper" layers) [53,54,145]. Lens-free imaging may be used to detect the aggregation of particles caused by bioreceptor–analyte interaction [55] (Figure 9). However, an exciting application is the direct, label-free classification of particles by lensless holography. Wu et al. [56] presented a lensless holography biosensor for classifying pollen and spores. As with many of these systems, a CNN was used for image reconstruction. In this work, another CNN was used to classify the particles, yielding > 94% accuracy.

Another work on the imaging classification of pollen utilizes multispectral imaging [58]. Again, a CNN was trained for classification, and a species-averaged accuracy of 96% was achieved for 35 plant species.

Artificial neural networks (ANNs) have also found grea<sup>t</sup> success in the developing field of digital staining. Hematoxylin and eosin (H&E) stain is the most common stain for histology [146]. However, the quality of tissue staining is subject to many factors that can affect the diagnosis. Digital staining is an alternative in which tissue sections are imaged unstained, and a trained model generates an image simulating stained tissue (Figure 10). Deep learning has been applied for digital staining on images acquired from a variety of methods including quantitative phase imaging [59], Fourier transformed infrared spectroscopy (FTIR) [52], and multi-modal multi-photon microscopy [57]. To overcome the issue of data scarcity and overfitting, researchers have frequently employed generative adversarial neural network (GAN) for medical imaging [147], which has shown promising results for digital staining model training [148]. Additionally, transfer learning has improved the model's generalizability to multiple domains [50].

Fluorescence-based imaging biosensors are also worthy of mention. Sagar et al. [51] presented a microglia classification based on fluorescence lifetime utilizing ANN.

The applications of imaging biosensors are extensive. Indeed, the scope is too large to analyze all papers in this review. However, of particular importance to imaging biosensors is the ANN, especially the CNN. This preference is expected since CNN has shown exceedingly good performance in a variety of image classification contexts [149,150].
