Region Proposal Algorithms to Train Support Vector Machines and Convolutional Neural Networks for Mass Detection in Mammograms

Contributors: Jaime Simarro Viana, Zohaib Salahuddin, Ahmed Gouda, Anindo Saha

Problem Statement: Extract candidate regions of interest (ROI) for mass detection in mammograms, that are subsequently to-be-used for training a classifier (eg. support vector machines, convolutional neural networks, etc.)

Dataset: INbreast Digital Mammographic Dataset - 115 cases (410 images). [Download - Credit: @wentaozhu]

Preprocessing (Contrast-Limited Adaptive Equalization)

Morphological Enhancement (Multi-Scale Morphological Sifting)

Unsupervised Segmentation (SLIC Superpixels)

Supervised Patch Extraction (Class-Weighted Sampling)

References

● Hang M. et al. (2019) "Multi-Scale Sifting for Mammographic Mass Detection and Segmentation", Biomedical Physics & Engineering Express, 5-2. DOI:10.1088/2057-1976/aafc07
● Radhakrishna A. et al. (2010) "SLIC Superpixels Compared to State-of-the-Art Superpixel Methods", IEEE TPAMI. DOI:10.1109/TPAMI.2012.120
● Moreira IC et al. (2012) "INbreast: Toward A Full-Field Digital Mammographic Database", Acad Radiol. 2012;19(2):236–248. DOI:10.1016/j.acra.2011.09.014