An Empirical Evaluation of Extreme Learning Machine Uncertainty Quantification for Automated Breast Cancer Detection

Early detection and diagnosis are the key factors in decreasing the breast cancer mortality rate in medical image analysis. A randomized learning technique called extreme learning machine (ELM) plays a vital role in learning the single hidden layer feed-forward network with fast learning speed and good generalization. The input weight and bias are randomly generated and fixed during the ELM training phase, and subsequently, the analytical procedure determines the output weight. The extreme learning machine’s learning ability is based on three uncertainty factors: the number of hidden nodes, an input weight initialization, and the type of activation function in the hidden layer. Various breast classification works have experimented with extreme learning machine techniques and did not investigate the following factors. This paper evaluates the extreme learning machine model’s performance with different configurations on the standard ultra-sound breast cancer dataset, BUSI. The proposed extreme learning machine configuration model experimented on original and filtered ultra-sound images. A fivefold stratified cross-validation scheme is applied here to enhance the model’s generalization performance. The proposed computer-aided diagnosis (CAD) model provides 100% accuracy with the best extreme learning machine configurations. Then, we compare the classification results of the proposed model with typical variants of extreme learning machines like Hybrid ELM (HELM), online-sequential ELM (OS-ELM), Weighted ELM, and complex ELM (CELM). The experimental results demonstrate that the proposed extreme learning machine model is superior to existing models, offering good generalization without any feature extraction or reduction method.