Meta-fusion network for tumor classification
Mayank Singh, Indu Saini, Neetu Sood · 2025
In the battle against cancer, Artificial Intelligence (AI) has a decisive role to play. AI has tremendous potential to facilitate initial-stage tumor detection. Countless Convolutional Neural Network (CNN) models exist, exclusively designed for pattern recognition. To implement specific tasks, one can inherit their intelligence using transfer learning and avoid training from scratch. Inheritance from the best-fit network can reduce the data requirement and training time. In this chapter, we proposed a meta-fusion network for tumor classification in ultrasound (US) images. Although the US has poor visible quality, it is the only harmless diagnostic imaging. We implemented the meta-fusion network by using the transfer learning and fusion layers. It is impractical to inherit from all the networks, so we have suggested inheriting modules from some of the most famous networks. The criteria for inclusion of a module of a network are unique. Nothing can match the efficiency of the activation map when it comes to analyzing any network. Using the activation map analysis and overlapping metrics, the fitness of a module was defined. The network was trained to classify tumors in breast and liver cancer. The dataset of 926 liver US and 780 breast US was divided into training, testing, and validation. Data augmentation, together with the validation set, ensured that the proposed network remained free of overfitting. Our meta-fusion network outperformed the original network in tumor classification. A detailed comparison of accuracy, sensitivity, and specificity with exclusive studies proposed recently for tumor classification is provided in the chapter. Meta-fusion network beats all of the networks in every aspect.