Effective Use of Deep Learning and Image Processing for Cancer Diagnosis

Abstract

The area of medical image processing obtains its significance with the requirement of precise and effective disease diagnosis over a short period. With manual processing becoming more complicated, stagnant and unfeasible with higher data size, there necessitates automatic processing that can transform contemporary medicine. Deep learning mechanisms can arrive at a higher rate of accuracy in processing and classifying images in comparison with human-level performance. Deep learning not only assist in selecting and extracting features but also possesses the potentiality of measuring predictive target audience and bestows prediction in a more action format to help doctors significantly. Unsupervised Deep Learning for cancer diagnosis is advantageous whenever the involvement of unlabeled data is huge. By bestowing unsupervised deep learning techniques to such unlabeled data, features of pixels that are superior compared to manually obtained features of pixels are said to be learned. Supervised Discriminating Deep Learning directly provides discriminating potentiality for cancer diagnosis purposes. Finally, hybrid deep learning for labeled and unlabeled data is specifically used for cancer diagnosis with a resource or poor pixel representations and hence early detection and diagnosis performed via bank features. Deep Neural Network, as the name implies includes several layers, emphasizing the complex non-linear relationships between the features present in the images, therefore contributing to higher accuracy. Deep Belief Network used in both supervised and unsupervised deep learning adopting greedy mechanism, maximizing the likelihood nature of detection and diagnosis at an early stage. Sequential event analysis is said to be performed by Recurrent Neural Network with the weights being shared across all neurons, contributing diagnosis accuracy. Certain fine-tuned learning parameters of consideration for better and precise learning are Interaction and Non-linear Rectified Activation function, Circumventing over-fitting via Dropout and Optimal Epoch Batch Normalization. In the last section, challenges about the application of deep learning for cancer diagnosis are discussed.