|ECR 2019 / C-2062|
|Representation learning approach to breast cancer diagnosis|
When compared with the base approach our new proposed method improves the classifier AUC from 0.826 to 0.978. The best approach uses the DL model for feature representation learning, in this case, the penultimate layer from the DL model, to feed a Random Forest classifier that enables, in average, to correctly classify 97.93% of the test data (BCDR-F03 dataset). The test in FFDM showed surprising results, giving us an accuracy of over 75% in all classifiers. This fact was significant since the existing samples of FFDM are not sufficiently representative to allow the creation of a dedicated classifier. Thus, the developed system is very promissory and could be used as a second opinion, in clinical workflows, producing acceptable results, although we consider that they have still not achieved the desired level of accuracy.
Finally, this is one more result that demonstrates that feature representation learning techniques are currently a good way to improve the performance of breast cancer CADx systems.
Thematically related posters
ECR 2019 / C-3045
Two techniques in comparison: Contrast-Enhanced Spectral Mammography versus Magnetic Resonance of the breast. Preliminary study
ECR 2019 / C-0211
Radioactive seed localisation and wire-guided localisation in excision of non-palpable breast cancer
ECR 2019 / C-1412
Added advantage of automated breast ultrasound in the detection of breast lesions in mammographically dense breast