Efficient machine learning algorithm for cancer detection using biomedical images

Abstract

The present work addresses a variety of challenges in the field of health informatics while using conventional machine learning and deep learning techniques for designing computer-assisted diagnostic systems. The work introduces new approaches to classify a complex dataset of breast cancer histopathological images, named as BreakHis . The preliminary part describes the types of handcrafted feature descriptors, different conventional machine learning approaches, deep learning techniques as well as the paradigm of transfer learning approaches along with their basic frameworks. Relevant studies in which the literature based on automated classification of images using different machine learning approaches are discussed in an organized manner to draw the important inferences. A convolutional neural network is proposed and trained from scratch to perform binary as well as multi-classification without considering the magnification factor of images. Further, the potential of transfer learning approach as a feature extractor and as a baseline model in magnification independent classification of breast cancer histopathological images is investigated using three pre-existing networks namely, VGG16, VGG19, and ResNet50. A comparative analysis is performed between conventional machine learning and transfer learning for multi-classification with due consideration of the magnification factor. In this context, the performance of a variety of soft computing algorithms is studied to determine the most efficient approach for the multi-classification of data. The role of layer-wise fine-tuning in the selection of the most effective depth of fine-tuning in AlexNet architecture at a different level of magnification is determined for binary as well as multi-classification of histological data. Eventually, the experimental results are demonstrated in the conclusive part which confirm the applicability of the designed approach in computer-aided diagnosis systems for cancer detection. newline