A new CNN architecture for efficient classification of ultrasound breast tumor images with activation map clustering based prediction validation

Effective ultrasound (US) analysis for preliminary breast tumor diagnosis is constrained due to the presence of complex echogenic patterns. Implementing pretrained models of convolutional neural networks (CNNs) which mostly focuses on natural images and using transfer learning seldom gives good results in medical domain. In this work, a CNN architecture, StepNet, with step-wise incremental convolution layers for each downsampled block was developed for classification of breast tumors as benign/malignant. To increase noise robustness and as an improvement over existing methodologies, neutrosophic preprocessing was performed, and the enhanced images were appended to the original image during training and data augmentation. The final layers' activation maps are clustered using fuzzy c-means clustering which qualify as a validation method for the prediction of StepNet. Using neutrosophic preprocessing alone had increased the validation accuracy from 0.84 to 0.93, while using neutrosophic preprocessing and augmentation had increased the accuracy to 0.98. StepNet has comparably less training and validation time than other state of the art architectures and methods and shows an increase in prediction accuracy even for challenging isoechoic and hypoechoic tumors.