Structured light imaging for breast-conserving surgery, part II: texture analysis and classification

Abstract

Subdiffuse spatial frequency domain imaging (sd-SFDI) data of 42 freshly excised, bread-loafed tumor resections from breast-conserving surgery (BCS) were evaluated using texture analysis and a machine learning framework for tissue classification. Resections contained 56 regions of interest (RoIs) determined by expert histopathological analysis. RoIs were coregistered with sd-SFDI data and sampled into ∼4 × 4 mm 2 subimage samples of confirmed and homogeneous histological categories. Sd-SFDI reflectance textures were analyzed using gray-level co-occurrence matrix pixel statistics, image primitives, and power spectral density curve parameters. Texture metrics exhibited statistical significance (p-value < 0.05) between three benign and three malignant tissue subtypes. Pairs of benign and malignant subtypes underwent texture-based, binary classification with correlation-based feature selection. Classification performance was evaluated using fivefold cross-validation and feature grid searching. Classification using subdiffuse, monochromatic reflectance (illumination spatial frequency of f x ¼ 1.37 mm −1 , optical wavelength of λ ¼ 490 nm) achieved accuracies ranging from 0.55 (95% CI: 0.41 to 0.69) to 0.95 (95% CI: 0.90 to 1.00) depending on the benign-malignant diagnosis pair. Texture analysis of sd-SFDI data maintains the spatial context within images, is free of light transport model assumptions, and may provide an alternative, computationally efficient approach for wide field-of-view (cm 2) BCS tumor margin assessment relative to pixel-based optical scatter or color properties alone.

Modeling and Synthesis of Breast Cancer Optical Property Signatures With Generative Models

“…It is during this transition, at n z = 10, ..., 40, where the introduction of finer details coincides with an improvement in classification performance for Fibrocystic Disease and Connective tissue, from ∼ 70% to ∼ 85% accuracy. In other words, performance improves as local structural variations surrounding a pixel is introduced and/or learned, to the point of allowing for individual identification, supporting parallel work that showed similar results on single-frequency, single-wavelength patch analysis [40]. It is important to note, however, that malignant tissue subtypes reach peak accuracy before higher-frequency texture is encoded in the keywords, suggesting the possibility of patient-specific spectral information that could be used in a case-by-case basis for margin delineation.…”

“…By employing bottleneck clamping, it is possible to observe that average spectral properties can be explained with few dimensions, and that texture presents as low-variance, high-dimensional fluctuations that are embedded within spectral information. Furthermore, it can be concluded that pixel-wise optical properties are sufficient for identifying malignant tissue, but that the inclusion of local textural information helps to uniquely identify categories with prominent textural features, such as Fibrocystic Disease and connective tissue, as long as inter-patient variability is compensated, supporting work that analyzes textural information exclusively [40]. Moreover, the dataset shows a detectable superposition between connective tissue and malignant tissue subtypes in feature space, suggesting that the presence of collagen and elastin in malignant growths, recently observed in multiphoton microscopy [36], could perhaps be measured macroscopically; further research is needed to ascertain if such presence of connective tissue could be quantified.…”

“…This suggests that specific tissue types could respond distinctively to spatial frequency modulation, while other diseases could be detected through this contrast improvement. For example, some specific pathologies reveal an inherent texture at high frequencies that cannot be observed in the low-frequency domain, further facilitating differentiability [39], [40]. This is notably visible in Figs.…”

Structured light imaging for breast-conserving surgery, part I: optical scatter and color analysis

“…Consequently, important textural information is omitted, which is considered in part II of this paper. 21 Textural information is known to be important in mammography. Its benefits occur at the pixel level because smaller regions can be included and allow the image analysis to retain its original spatial resolution.…”

Biophotonic technologies for assessment of breast tumor surgical margins—A review

“…Maloney et al used optical scattering analysis on SFDI data of 70 BCS specimens and demonstrated an extensive categorization of benign tissues into four types and malignant tissue into six types [85]. As scatter imaging involves long processing times and also susceptible to other sources of errors, the team investigated color analysis [85] and texture analysis [86] individually as a surrogate for the former. Both color and texture analyses enabled faster processing times but at the compromise of the categorization.…”