preview

Essay On Automatic Diagnosis For Cancer

Good Essays
Open Document

Automatic diagnosis for cancer is a ongoing topic of research. Previous research have been conducted on breast cancer diagnosis. Doyle et al. [9] investigated the classification of breast cancer histopathology from textual, nuclear architectural features and images. Classification of benign and malignant histopathological im- ages based on 10 syntactic structure features, extracted after the Gabor filter had been applied was used by van Diest et al.[10]. The usage of Support Vector Ma- chines (SVMs) on relevant numerical features extracted from histopathological images in a semi-automatic fashion has also yielded very promising results[11].
George et al.[12] developed an intelligent remote detection and diagnosis system for breast cancer …show more content…

tasks[15][16][17].
However, deep CNNs often have a number of parameters so large that it cannot reasonably be trained without a very large dataset. Medical imaging datasets are often not sufficiently large enough to train a deep CNN model from scratch adequately. Thus, the usage of transfer learning in medical imaging has been explored[18]. Transfer learning seeks to transfer knowledge between large source and target domains[5]. For CNNs, this is often done by pre-training a CNN model with a large dataset, then re-training the model with the smaller target dataset[19]. Yosinski et al.[20] showed that features in deep neural networks, in particular those learnt in bottom layers of a CNN are transferable, and fine- tuning the entire CNN model with the target dataset improves the performance of model. Shin et al.[18] demonstrated that the performance of deep CNNs on several CADe problems can consistently improved by transfer learning from the large scale natural image dataset, ImageNet.
3
3.1
Approach
Dataset
The dataset used for training and evaluation is the benchmark BreakHis dataset[7].
It contains 7909 breast cancer histopathology images acquired from 82 patients.
These images are labeled either ”Benign” or ”Malignant” and can be of four different resolutions (40X, 100X, 200X, 400X). The dataset is subsequently sep- arated into five folds. Different samples from the same

Get Access
Related
Get Access