Deep Learning Models for Histopatological Images Classification

EasyChair Preprint 331

Abstract

This study aimed to compare five deep learning models for classifying Squamous Cell Carcinoma (SCC) tumor of the head and neck cancer from histopatplogical images. The five deep learning models were pre-trained convolutional neural networks (CNNs) included googleNet, Inception-v3, ResNet-50, ResNet-101 and Inceptionresnet-v2. These pre-trained CNNs were used to build fine-tuning networks with transfer learning. The building transfer learning networks replace the last three layers of pre-trained CNNs which configured for 1000 classes by new layers for binary classes and then fine-tune these layers on the Head and Neck Squamous Cell Carcinoma (HNSCC) tumor images. A total of 1,424 histopatological images of head and neck cancer were used to detect the tumor cells in sections. The transfer learning networks were compared in terms of standard performance.  Although the number of images was insufficient, the results were shown good accuracy among different models. A highly successful classification has been achieved by the ResNet-50 model with accuracy rate was %98.95. But ResNet-101, googleNet and Inception-v3 performed classification with accepted accuracy rates of %97.89, %97.19 and 94.04%, respectively.

Keyphrases: Deep Convolution Neural Network, Head and neck cancer, Standard Performance, image classification

@booklet{EasyChair:331,
  author    = {Maisun Al Zorgani and Hassan Ugail},
  title     = {Deep Learning Models  for Histopatological Images Classification },
  doi       = {10.29007/1v9h},
  howpublished = {EasyChair Preprint 331},
  year      = {EasyChair, 2018}}