Weakly supervised learning for breast cancer detection in ultrasound images.
Abstract
Large amounts of well data is required to achieve a good model performance. It is very
costly and the available time for manual annotation is very limited to acquire more labelled
data. In Uganda there are limited number of radiologists whose expertise and experience
also differ which brings about costly and many times wrong diagnosis. To solve this chal-
lenge we employed weakly supervised learning techniques to detect the presence of breast
cancer in ultrasound images. This is less costly and automated method for annotation of
images to obtain image-level labels with a bounding box. The deep learning model was
trained on a mix of local and opensource dataset with part of it already labelled by the
experts with only image level-labels and another part with ground truth annotations of
bounding boxes. The yolov4-tiny framework results were very sounding and optimal with
a mAP of 91.5%.
The complete data-set was the back bone in the develop model for recognition, classifica-
tion and localization of Breast Cancer in Ultrasound images.