HISTOPATHOLOGY IMAGE CLASSIFICATION USING HYBRID PARALLEL STRUCTURED DEEP-CNN MODELS

Abstract

The healthcare industry is one of the many out there that could majorly benefit from advancement in the technology it utilizes. Artificial intelligence (AI) technologies are especially integral and specifically deep learning (DL); a highly useful data-driven technology. It is applied in a variety of different methods but it mainly depends on the structure of the available data. However, with varying applications, this technology produces data in different contexts with particular connotations. Reports which are the images of scans play a great role in identifying the existence of the disease in a patient. Further, the automation in processing these images using technology like CNN-based models makes it highly efficient in reducing human errors otherwise resulting in large data. Hence this study presents a hybrid deep learning architecture to classify the histopathology images to identify the presence of cancer in a patient. Further, the proposed models are parallelized using the TensorFlow-GPU framework to accelerate the training of these deep CNN (Convolution Neural Networks) architectures. This study uses the transfer learning technique during training and early stopping criteria are used to avoid overfitting during the training phase. these models use LSTM parallel layer imposed in the model to experiment with four considered architectures such as MobileNet, VGG16, and ResNet with 101 and 152 layers. The experimental results produced by these hybrid models show that the capability of Hybrid ResNet101 and Hybrid ResNet152 architectures are highly suitable with an accuracy of 90% and 92%. Finally, this study concludes that the proposed Hybrid ResNet-152 architecture is highly efficient in classifying the histopathology images. The proposed study has conducted a well-focused and detailed experimental study which will further help researchers to understand the deep CNN architectures to be applied in application development.