Horizontal Cells and Retinal Ganglion Receptive Fields (Vision Series Part 3)

Visual information travels through a well-defined pathway from the eye to the brain. The optic nerve carries axons from retinal ganglion cells to the optic chiasm, where nasal retinal fibers cross to the contralateral side while temporal fibers remain ipsilateral. The suprachiasmatic nucleus receives circadian rhythm information from light-sensitive ganglion cells. Axons project to the lateral geniculate nucleus (LGN) of the thalamus, which contains six layers organized into magnocellular (layers 1-2) and parvocellular (layers 3-6) divisions. Magnocellular layers receive input from M-type ganglion cells in the periphery and process motion, while parvocellular layers receive input from P-type ganglion cells in the fovea and handle detail and color. Both LGN and retina maintain retinotopic maps, meaning specific retinal locations project to specific brain locations.

Retinal ganglion cells are glutamatergic output neurons that project to various brain structures. About 1% are intrinsically photosensitive (project to midbrain for pupillary reflex). Most project to the lateral geniculate nucleus (LGN) of the thalamus, which then projects to the primary visual cortex. The optic nerve must pierce the retina, creating the blind spot where no photoreceptors exist.

Visual information travels from the retina through several key structures: the optic nerve exits at the blind spot, converges at the optic chiasm where nasal retinal fibers cross, then continues as the optic tract to the lateral geniculate nucleus (LGN) in the thalamus. Approximately 90% of retinal ganglion cell axons project to the LGN, which serves as the primary relay station before reaching the primary visual cortex (V1). This pathway demonstrates how visual information is systematically routed from the eye to higher brain centers.

Visual information travels through retina-geniculate-striate pathways: retinal ganglion cell axons form the optic nerve, passing through lateral geniculate nuclei (thalamus) to the primary visual cortex in occipital lobes. A critical feature is crossed projection—left visual field inputs go to right hemisphere, and vice versa. The LGN contains parvocellular layers (small cells for color and detail) and magnocellular layers (large cells for motion). The visual cortex contains vertical columns corresponding to specific retinal locations. Processing hierarchically progresses from primary cortex → secondary cortex (prestriate and inferotemporal) → visual association cortex (posterior parietal and distributed areas), enabling progressive analysis from basic features to complex interpretations.

Visual information travels from the retina through the optic nerve to the lateral geniculate nucleus (LGN). At the optic chiasm, nasal retinal axons cross to the opposite brain hemisphere while temporal axons remain on the same side. The LGN, located within the thalamus, receives organized input from both eyes. Each eye sends axons to both LGNs based on retinal origin: right eye nasal retina projects to left LGN layers 2,3,5; right eye temporal retina projects to right LGN layers 1,4,6. The LGN contains six distinct layers that process different aspects of visual information, serving as a critical relay station before transmitting signals to the visual cortex.

Retinal ganglion cells and V1 neurons have receptive fields with center-surround organization. In an on-center cell, stimulation in the center depolarizes the cell while stimulation in the surround hyperpolarizes it. This organization enhances contrast detection and edge perception. When a stimulus is placed in the center, the cell fires; when placed in the surround, the cell is inhibited. This center-surround organization is preserved throughout the visual pathway and is fundamental for detecting edges and boundaries in visual scenes.

A receptive field is the area over which a neuron receives information. At each visual system level, neurons receive input only from a small window of previous layer neurons. Ganglion cells have center-surround receptive fields: on-center cells increase firing with central light and decrease with peripheral light; off-center cells do the opposite. When light fills the entire receptive field, effects cancel out. This organization allows extraction of edges and contrast, which is fundamental for visual processing. The LGN has six sub-layers receiving segregated inputs from both eyes. The magnocellular pathway (large cells) goes to layers 1-2, carrying motion and contrast information. The parvocellular pathway (small cells) goes to layers 3-6, carrying color and fine detail. Contralateral inputs go to layers 1, 4, 6; ipsilateral inputs go to layers 2, 3, 5. LGN cells have circular receptive fields with stronger surround influences than ganglion cells, amplifying contrast information before reaching the cortex.

Receptive fields define what stimuli activate a neuron. Retinal ganglion cells have center-surround receptive fields: on-center cells fire to light in the center with dark surround; off-center cells fire to dark center with light surround. Bipolar cells (on and off types) and horizontal cells create this organization through excitatory and inhibitory connections. Horizontal cells provide lateral inhibition, increasing contrast sensitivity by suppressing neighboring photoreceptors. This circuit motif enhances edge detection and is fundamental to sensory processing across modalities.

Visual perception emerges from hierarchical processing in the visual cortex. Photoreceptors in the retina transduce light into electrical signals, which travel through the thalamus to the primary visual cortex (V1). Neurons in V1 possess orientation-selective receptive fields—they fire maximally when stimuli match their preferred orientation but respond minimally to adjacent orientations. This selectivity enables edge detection. LGN neurons with simple on-center/off-surround receptive fields combine to create V1's complex orientation-selective cells. Beyond V1, parallel pathways process spatial location (dorsal stream) versus object identity (ventral stream), with higher cortical areas developing increasingly abstract representations of visual stimuli.

V1 neurons show multiple receptive field properties: orientation selectivity, direction selectivity, spatial frequency selectivity, and binocular convergence. While LGN cells are monocular, most V1 cells above layer 4CL receive input from both eyes. ON/OFF convergence becomes complete in complex cells. Hubel and Wiesel discovered ocular dominance columns using transneuronal transport labeling, appearing as alternating stripes. Subsequent work revealed orientation selectivity columns perpendicular to ocular dominance columns. Multiple models (ice cube, radio, swirl) attempted to explain this organization, with optical recording supporting the swirl model as the actual cortical arrangement. The cytochrome oxidase patches, always located at ocular dominance column centers, serve as metabolic activity markers.

The Difference of Gaussians (DoG) is an edge detection technique that applies two Gaussian blurs with different standard deviations to an image and subtracts them, creating a band-pass filter that isolates high-contrast edges; this method, originally developed in 2012 by Windmiller and colleagues, can be enhanced with extended parameters for brightness preservation, noise reduction through edge-aligned line integral convolution, and various artistic stylizations including charcoal, pastel, watercolor, and cross-hatching effects.

Band-pass filters can be created by subtracting a large blur from a small blur (Difference of Gaussians or DOG). Small blurs remove only very high frequencies; large blurs remove highs, mids, and lows. Subtracting cancels lows (present in both) and removes highs (removed by both), leaving only the middle band. DOG approximates the Laplacian of Gaussian (LOG), which has a Mexican hat shape. The LOG is similar to combined blur-sharpen kernels but produces slightly darker patterns. Using two separate blurs allows easier experimentation with different blur types and sizes.

The Difference of Gaussians (DoG) detector achieves scale invariance by blurring an image with multiple Gaussian kernels at different sigma values (e.g., 1, 2, 4, 8 pixels) and computing the difference between consecutive blurred images. For each pixel, the algorithm searches for maxima not only in the spatial X and Y directions but also across the scale pyramid. This ensures that the same physical feature will produce a response at the appropriate scale level regardless of whether the image was captured from closer or farther away. The DoG detector then performs non-maximum suppression in both space and scale, followed by edge elimination to retain only corner-like features.

The Difference of Gaussians (DoG) is a preprocessing technique that enhances high-frequency details in images by subtracting one blurred version of the image from another. Unlike the Sobel filter which is designed for feature extraction rather than aesthetic results, DoG produces visually pleasing edges that better suit artistic applications. When combined with Sobel filtering for edge direction, DoG improves both the visual quality and reliability of edge detection in ASCII art shaders.

Laplacian of Gaussian (LoG) is the second derivative of Gaussian, acting as a blob detector. Difference of Gaussians (DoG) approximates LoG by subtracting two Gaussian-blurred images with different sigma values. DoG is computationally cheaper while producing similar results. DoG images are computed at multiple scales, and differences between adjacent scales highlight scale-invariant features. This forms the basis for scale space extrema detection in SIFT.