Abstract 90P
Background
Breast cancer is the most common female cancer worldwide. Breast screening programmes have increased detection of the earliest stages; however, mammography suffers from high false positivity and is labour intensive for radiologists. Recognition of breast abnormalities, which mainly consists of mass and micro-calcification, plays a critical role in the detection of breast cancer.
Methods
We used the open accessed mini-MIAS dataset. 100 abnormal and 100 healthy images were selected randomly. The abnormal class contains six sub-types: asymmetry, distortion, calcification, ill-defined mass, circumscribed mass, and spiculated mass. A preprocessing technique was carried out by (i) removing additive and multiplicative noises and (ii) removing the background and pectoral muscle regions. For the classifier, we chose to use the popular AlexNet model. In order to improve its performance, we integrated the following improvements by: (i) reducing the kernel size of the first conv layer from 11x11 to 7x7, so our network can learn subtle feature information. (ii) utilizing ELU function to replace the traditional ReLU function. (iii) optimizing essential hyperparamters, including dropout ratio and learning rate. (iv) using data augmentation (rotation, translation, scaling, and contrast-adjusting) techniques to enhance the size of training set. Some improvements are inspired from Ref. [Liu Z, et al., JFE, 2020 & Clevert DA, et al., Arxiv, 2015].
Results
The average of 20 runs of 10-fold cross validation showed this improved AlexNet deep neural network model can achieve a sensitivity of 93.80%, a specificity of 94.10%, a precision of 94.08%, an accuracy of 93.95%, and F1 score of 93.94%. This proposed method gives better performance than several state-of-the-art approaches, inlucing [Guo Z, et al., IEEE ROMAN, 2018, & Liu F, et al., LNCS 2019].
Conclusions
This proposed AI method is effective in detecting abnormal breasts. In the future, we shall try to extend our AI system into multi-class diagnosis systems together with the function of lesion region prediction. We expect the callback rate of abnormal breast diagnosis can be effectively reduced.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.