Deep-stacked autoencoder for medical image classification
J. Anitha, S. Akila Agnes, S. Immanuel Alex Pandian, Malin Bruntha · Institution of Engineering and Technology eBooks · 2022
Image classification is the most important process in the computer-aided diagnosis system. Feature learning is the major challenge in the classification task where the representation features are refined from the high-dimensional input data. Autoencoder (AE) is a self-supervised neural network that maps the input data with the target output data through the sequence of encoder and decoder layers. It automatically learns the abstract-level representation features with the bottleneck feature of encoder-decoder based neural networks. The encoder layers are trained on various levels in extracting the essential features to represent the image. At last, all levels of encoders are stacked together to form a stacked network. The decoder part is replaced by the final softmax layer for classification. The complete stacked network is fine-tuned for classification in a supervised fashion. In this chapter, a stacked AE model is developed to classify the medical images, particularly skin cancer images, into benign or malignant. From the experimental results, it is noted that the classification accuracy of a stacked AE with fine-tuning is higher than that of a stacked AE without fine-tuning. The experimental results also confirm that the stacked AE model with fine-tuning provides improved results as compared with other classification methods.