Using convolutional neural networks for the classification of breast cancer images

TL;DR

This study evaluates various convolutional neural networks and hardware accelerators for classifying breast cancer tissue images, demonstrating that model complexity and transfer learning impact performance and training efficiency.

Contribution

It compares twelve CNN architectures and hardware devices on large public datasets, analyzing the effects of transfer learning and hyperparameter tuning in breast cancer image classification.

Findings

Hardware acceleration significantly reduces training time.

Increasing model depth improves accuracy but increases training time.

Transfer learning performs worse than full retraining in this context.

Abstract

An important part of breast cancer staging is the assessment of the sentinel axillary node for early signs of tumor spreading. However, this assessment by pathologists is not always easy and retrospective surveys often requalify the status of a high proportion of sentinel nodes. Convolutional Neural Networks (CNNs) are a class of deep learning algorithms that have shown excellent performances in the most challenging visual classification tasks, with numerous applications in medical imaging. In this study I compare twelve different CNNs and different hardware acceleration devices for the detection of breast cancer from microscopic images of breast cancer tissue. Convolutional models are trained and tested on two public datasets. The first one is composed of more than 300,000 images of sentinel lymph node tissue from breast cancer patients, while the second one has more than 220,000 images from inductive breast carcinoma tissue, one of the most common forms of breast cancer. Four different hardware acceleration cards were used, with an off-the-shelf deep learning framework. The impact of transfer learning and hyperparameters fine-tuning are tested. Hardware acceleration device performance can improve training time by a factor of five to twelve, depending on the model used. On the other hand, increasing convolutional depth will augment the training time by a factor of four to six times, depending on the acceleration device used. Increasing the depth and the complexity of the model generally improves performance, but the relationship is not linear and also depends on the architecture of the model. The performance of transfer learning is always worse compared to a complete retraining of the model. Fine-tuning the hyperparameters of the model improves the results, with the best model showing a performance comparable to state-of-the-art models.