“Understanding the effects of image filtration and quantization in extracting pathologically relevant texture features"

Radiomics is a vital tool in modern medicine, enhancing precision in diagnosis and treatment by extracting a diverse range of disease-characteristic features from medical images through advanced data analysis algorithms. This approach provides insights into pathology by analyzing image texture, including characterizing pixel brightness distributions, spatial patterns, and structures. As a partial-angle 3D X-ray technique used in mammography, Digital Breast Tomosynthesis (DBT) reduces structural noise resulting from overlap, leading to improved detection and characterization of breast lesions. This makes DBT valuable for texture analysis and enhances clinical performance in breast cancer detection. Despite these advantages, distinguishing tissue types in DBT images can be challenging.

We examined the role of image textures and assessed how various image processing methods affect these features. The use of simulated or virtual imaging offers ground truth for comparison and evaluation of changes in quantitative estimates with varying imaging parameters. We used simulated DBT images with accurate X-ray attenuation properties for cancer and other breast tissues in a complex digital phantom. We then evaluated the effects of several commonly used image processing techniques—median filtering, Gaussian blurring, and unsharp masking—on quantitative changes in texture and lesion detection. Median filtering reduced noise while preserving edges, Gaussian blurring smoothed the image and reduced high-frequency noise, and unsharp masking enhanced edge definition and contrast.

Analysis of adipose and fibroglandular tissue regions at different quantization levels showed that median filtering and unsharp masking, combined with higher quantization levels, provided a more detailed texture representation and improved lesion detection accuracy. While lower quantization levels were adequate for basic analysis, higher levels enhanced the differentiation between lesions and surrounding tissues. These findings highlight the importance of selecting appropriate quantization levels and processing techniques to enhance biomarker classification and improve breast cancer lesion detection.