A cross-cutting approach for tracking architectural distortion locii on digital breast tomosynthesis slices. (April 2019)

• Record Type: Journal Article
• Authors: de Oliveira, Helder C.R.; Mencattini, Arianna; Casti, Paola; Catani, Juliana H.; de Barros, Nestor; Gonzaga, Adilson; Martinelli, Eugenio; da Costa Vieira, Marcelo A.
• Is Part Of: Biomedical signal processing and control. Volume 50(2019)
• Page Start: 92
• Page End: 102
• Publication Date: 2019-04
• Subjects: Architectural distortion -- Digital breast tomosynthesis -- Breast cancer -- Computer aided detection -- Gabor filter -- Cell tracking; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2019.01.001
• Languages: English
• ISSNs: 1746-8094

Abstract

Graphical abstract: Highlights: Automatic localization of architectural distortion candidates in digital breast tomosynthesis slices by Gaussian curvature computation and Gabor filtering approaches. Automatic selection of candidate AD paths along the slices using cell-tracking algorithms. Unsupervised slice-by-slice feature extraction in the regions of interest around the track. Discrimination model between true and false AD tracks. Abstract: Background and objective: Full-field digital mammography (FFDM) is the predominant breast cancer screening exam used. However, with the emergence of digital breast tomosynthesis (DBT) the radiologists could improve early recognition of breast cancer signs. In this scenario, the detection of architectural distortion (AD) is still a challenging task. ADs are very subtle contraction of the breast parenchyma that could represent the earliest manifestation of cancer, assessing at present 50% of missed cases. Methods: This paper proposes a new paradigm to detect AD in DBT exams by a cross-cutting approach exploiting the 3-dimensionality of the imaging modality. After locating AD candidates in each DBT slice, the suspicious spots are tracked in cross-slice direction and then characterized in terms of neighboring texture. In this approach, which mimics radiologist's scrolling down over zoomed slices, we reduce the amount of uninformative signs collected in DBT exams by preserving the large variability of AD appearance. Results: Using 37 sets of DBT slices containing at least one AD locus indicated by a radiologist, the proposed methodology reaches an AUC of 0.84, with only one false negative exam at sensitivity of 0.9. Conclusions: The results show that the proposed algorithm can be a promising tool for the automatic detection of AD locii. Future work will address the extension of the dataset of DBT slices as well the improvement of algorithm performance toward the application in the clinical practice.