LARSE: level-based associated residual network with squeeze-and-excitation for breast cancer detection and classification

Breast cancer is considered as a serious disease causing a high mortality rate amongst women. In recent years, computer aided diagnosis (CAD) techniques have the radiologists to make proper decisions on mammograms more accurately. The existing CAD method may not contribute significant results for the early identification of breast mass especially at stages 1 and 2. This work introduces a level-based associated residual network with squeeze-and-excitation (LARSE) block for breast cancer classification. Initially, the input image undergoes pre-processing using the contrast limited adaptive histogram equalisation (CLAHE) model. Then, the feature extraction process is done by utilising a dual ResNet-based feature extraction model, LARSE. The LARSE model is used for multilevel breast cancer classification based on BI.RADS categories, tested on CBIS.DDSM and INbreast mammogram datasets. The LARSE model achieved an accuracy of 96.9% (±0.5%), sensitivity of 98% (±0.4%), specificity of 94% (±0.6%), and F1.score of 96% (±0.3%).