Automatic breast vascular network feature extraction in 3D photoacoustic tomography images: preliminary results on cancer-affected and healthy breasts

Background: According to the European cancer information system, breast cancer is the most prevalent female cancer in Europe, and it has the highest mortality rate. Thanks to screening programs, the mortality rate has been decreasing in the last decades. The gold standard for breast cancer screening is to perform an X-ray mammography every 2 years for women above the age of 50 years. Therefore, women are periodically exposed to a certain amount of radiation, limiting the screening frequency. This limitation could be overcome by using non-ionizing imaging techniques, increasing the probability of detecting the presence of cancer. A possible solution is given by photoacoustic tomography (PAT), which combines optical absorption contrast with the imaging depth of ultrasound-based techniques. In the last years, photoacoustic imaging has been shown to be a promising modality to non-invasively study breast cancer due to its ability to depict the tumor-associated vasculature. The latter is the result of angiogenesis, which is the process of creating new blood vessels from preexisting ones. Although it is well known that angiogenesis influences the local vasculature around the tumor site, its impact on the global vascular network of the host organ is still unclear. Objective: This work aims to design and test an automated image analysis pipeline for extracting quantitative features from vascular networks in breast PAT images, in order to evaluate differences in the global vascular network between cancer-affected and healthy breasts. Dataset and methods: Twenty-four 3D breast volumes acquired with the University of Twente’s Photoacoustic Mammoscope 3 (PAM3) were analyzed. The dataset included 14 healthy and 10 cancer-affected breast images. The image analysis pipeline developed for this study involved multiple steps: preprocessing, with the application of Hessian-based Frangi filtering and adaptive intensity modulation for enhancement of blood vessels; vascular network extraction, performing an intensity and depth adaptive segmentation followed by skeletonization using Lee’s algorithm. Finally, several architectural (e.g., vessels’ diameters and densities), tortuosity (e.g., distance metric), and intensity-based features (e.g., entropy map) were extracted. Analysis of variance (ANOVA) and multivariate analysis of variance (MANOVA) were performed to determine the effectiveness of the extracted features in distinguishing between healthy and cancer-affected breasts. Results: Cancer-affected breasts showed a higher mean value of the 75th percentile of vessels’ length-normalized diameters than healthy breasts (ANOVA P<0.05). The MANOVA dimension of the group mean values was equal to 1 (P<0.05) with the following most important features in the differentiation: i) 25th percentile of distance metric values, ii) 95th percentile of vessel intensities, and iii) density of main vessels. Conclusion: Consistent with prior studies, these results suggest that the presence of cancer might influence the vascular network of the host organ. To support this assertion, a set of quantitative features related to the presence of cancer, including architectural, tortuosity, and intensity-based features, has been identified. However, these results need to be confirmed by further studies, which will include a larger dataset. In conclusion, the study achieved promising results as the first feasibility study of a computer-aided system based on a non-invasive technology that does not use ionizing radiation.