Laplacian of Gaussian (LoG) Filter

A computationally efficient approximation of the LoG filter achieved by subtracting two Gaussian-blurred versions of an image at different scales. Instead of applying a Laplacian after smoothing, DoG convolves the image with two Gaussians of variance σ₁ and σ₂ and takes the difference.

• Scale Normalization: Requires multiplication by σ² to achieve true scale invariance
• Physiological Plausibility: Models the center-surround receptive fields in retinal ganglion cells
• SIFT Integration: Forms the scale-space extrema detection backbone of the Scale-Invariant Feature Transform
• Radiomic Application: Used to detect blob-like tumor structures at multiple anatomical scales simultaneously

A formal framework for handling image structures at multiple scales by embedding the original image into a one-parameter family of smoothed images. The Gaussian kernel is the unique linear scale-space kernel proven by the non-enhancement of local extrema property.

• Scale Parameter (σ): Continuous parameter controlling the level of observation detail
• Causality Structure: Coarser scales must be simplifications of finer scales—no new structure is created
• Differential Invariants: Combinations of Gaussian derivatives that remain constant under rotation and illumination changes
• Deep Structure: Tracks how zero-crossings and blobs evolve, merge, and annihilate across scales

The method for localizing edges by identifying pixels where the LoG response changes sign, indicating a transition from positive to negative second derivative. This corresponds to the inflection point of intensity change.

• Sub-Pixel Accuracy: Achieved through linear interpolation of the zero-crossing surface
• Closed Contours: Guarantees continuous, unbroken edge maps—a key advantage over gradient-based methods
• False Positive Suppression: Requires gradient magnitude thresholding to eliminate zero-crossings from noise
• Marr-Hildreth Edge Detector: The canonical algorithm combining LoG filtering with zero-crossing localization

The LoG function is mathematically identical to the Ricker wavelet, commonly called the Mexican hat due to its characteristic shape—a central positive lobe surrounded by a negative annular region.

• Continuous Wavelet Transform: The LoG serves as an admissible mother wavelet for time-frequency analysis
• Band-Pass Behavior: Acts as a spatial band-pass filter, suppressing both low-frequency background and high-frequency noise
• Normalization Factor: The 2D LoG kernel is normalized by -1/(πσ⁴) to ensure zero DC response
• Separability Limitation: Unlike the Gaussian, the LoG kernel is not separable but can be decomposed via DoG approximation

The primary application of LoG filtering in radiomics: identifying circular or ellipsoidal regions of homogeneous intensity that differ from their surroundings. The LoG response is maximal when the zero-crossings of the Laplacian align with the boundary of a blob.

• Characteristic Scale: The σ at which the LoG response peaks corresponds to the blob radius r = σ√2
• Automatic Scale Selection: Searching for local maxima in the 3D (x, y, σ) scale-space volume
• Tumor Nodule Detection: Applied in lung CT screening to identify candidate pulmonary nodules
• Lesion Boundary Initialization: Provides seed points for active contour and level-set segmentation algorithms

The LoG is a specific member of the N-jet family: the set of all Gaussian partial derivatives up to order N. These filters form a complete, stable basis for local image structure description.

• Steerability: Linear combinations of Gaussian derivatives can synthesize filters at any orientation
• Hessian Matrix: The second-order derivatives (Lxx, Lyy, Lxy) compute principal curvatures for vesselness measures
• Frangi Vesselness: Uses Hessian eigenvalues to enhance tubular structures in angiography
• IBSI Compliance: Gaussian derivative features are standardized in the Image Biomarker Standardisation Initiative