Authors:
Clécio Silva
;
Salomão Mafalda
;
Emili Bezerra
;
Gustavo Oliveira de Castro
;
Paulo Santos Júnior
and
Ana Alvarez
Affiliation:
PAVIC Laboratory, University of Acre (UFAC), Rio Branco, Acre, Brazil
Keyword(s):
Breast Cancer, Semantic Segmentation, Neural Networks, Deep Learning.
Abstract:
Breast cancer is a problem that affects thousands of people every year, early diagnosis is important for the
treatment of this disease. Deep learning methods shows impressive results in identification and segmentation
of breast cancer task. This paper evaluates the impact of input size images on three semantic segmentation
architectures applied to breast tumour ultrasound, in U-net, SegNet and DeepLabV3+. In order to (comprehensively) evaluate each architecture, 5-fold cross validation was carried out, thus reducing the impact of
variations in validation and training sets. In addition, the performance of the analyzed architectures was measured using the IoU and Dice metrics. The results showed that the DeepLabV3+ architecture performed better
than the others architectures in segmenting breast tumours, achieving an IoU of 0.70 and Dice of 0.73, with
the input dimension of the images being 128×128.