<p>Computer-aided-detection (CAD) systems have been evolved to aid radiologists in identifying skeptical sores in mammogram. Deep learning algorithms have recently taken the place to boost the probability of discerning malformation at an early stage with the intention of keeping away from irrelevant biopsies and minimize the rate of mortality. In recent years, deep learning is a recent research technique aimed at classifying mammograms gene expression. This paper is to develop adaptive kernel attention U-Net and Bi-direction backward propagated ResNet-based (AKA-BBP) mammogram classification and detection. In the initial phase, the raw set of images is gathered from the Mini_MIAS dataset. The projected method is proceeding with four stages pre-processing, segmentation, feature selection and classification. In the pre-processing phase, the adaptive kernel Gaussian probability density-based preprocessing model is considered. After that regions of interest are achievedwith the consideration of spatial and channel attention-based U-Net for breast segmentation. The segmented features will be utilized for selecting the best feature subset. Finally in the classification stage, backward propagated ResNet neural network-based classifier is used. In the classification state occurrence of error in cases are handled using backward propagation. The presentation of the proposed AKA-BBP methodology is evaluated in PSNR, precision, recall and f-measure. To validate the projected method, AKA-BBP, it is compared with self supervised learning (SSL) method, haze reduced adaptive technique and chaotic crow search optimization algorithm (HRAT-CCSOA, and fusion of hybrid deep features (FHDF). The quantitatively analyzed results show that the AKA-BBP method improved 19% of precision, 12% of recall and 38% of PSNR with preprocessing. In addition, the numerical outcome of the AKA-BBP method provides a better improvement in 16% F-score compared to conventional techniques.</p>