A Study of the Generalisability of CNNs for Disease Prediction

Convolutional neural networks (CNNs) have proven to be effective for disease prediction from images, namely, CT scans. However, previous research studies the application of CNNs for disease prediction separately for each disease. The aim of this study is to investigative the generalisability of CNNs for disease prediction. The study firstly examines the performance of two neural networks, namely, AlexNet and ResNet34 for disease prediction across three diseases, namely, brain tumour diagnosis, lung cancer detection and skin cancer diagnosis. The study also investigates an incremental neural network approach (INNA) for learning in neural networks when used for disease prediction. The INNA divides the dataset into easy, medium and hard using a difficulty estimation techniqu The performance of these CNNs and the INNA are evaluated independently for each of the three diseases examined, as well as across all three diseases. The study revealed that when evaluated separately for each disease AlexNet performed the best for brain tumour diagnosis and ResNet34 for lung cancer detection and skin cancer diagnosis. When evaluated across all three datasets for the different diseases, ResNet34 outperformed AlexNet and both CNNs with INNA. While the INNA did not outperform AlexNet or ResNet34, it had a lower computational cost and outperformed state of the art approaches for lung cancer detection.