Improved Watershed Segmentation and DualNet Deep Learning Classifiers for Breast Cancer Classification

Breast cancer is a deadly, noncommunicable disease that affects women worldwide. Early detection is crucial in providing effective treatment and improving survival rates. Towards that, a novel DualNet-deep-learning-classifier model for efficient and accurate breast cancer detection is proposed. The proposed model includes five major phases: “pre-processing, segmentation, feature extraction, feature selection, and breast cancer detection”. The pre-processing phase involves noise removal via Wiener filtering and image contrast enhancement via a contrast stretching approach. Then, from the pre-processed mammogram images, the ROI region is identified using the new gradient-based watershed segmentation approach. Subsequently, from the identified ROI regions the texture [grey level run-length matrix (GLRLM), multi-threshold rotation invariant LBP (MT-RILBP) (proposed)], color (color correlogram), and shape features (Zernike moment) are extracted; and among the extracted features, the optimal features are chosen using a hybrid optimization model-FlyBird optimization algorithm (FBO), which incorporates both the “fruit fly optimization algorithm and bird mating optimizer”. The breast cancer classification phase uses a DualNet-deep-learning-classifiers approach that includes “long short-term memory networks (LSTM), convolutional spiking neural networks (CSNN), and a new optimized autoencoder (OptAuto)”. The LSTM and CSNN are trained using the identified optimal features. The outcome from LSTM and CSNN is fed as input to OptAuto, wherein the outcome regarding the presence/absence of breast cancer is identified. Moreover, the weight function of the autoencoder is tuned using the new FBO. The proposed model is evaluated in terms of “sensitivity, accuracy, specificity, precision, TPR, FPR, TNR, F1-score, and recall”. Overall, the proposed model holds promise for accurate and efficient breast cancer detection, with potential for future clinical applications.