Automatic classification of ovarian cancer types from cytological images using deep convolutional neural networks

Miao Wu, Chuanbo Yan, Huiqiang Liu, Qian LiuBioscience 2018 · DOI

TL;DR

This paper introduces a deep learning approach using a DCNN based on AlexNet to automatically classify four subtypes of ovarian cancer from cytological images. By augmenting the training data through image enhancement and rotation, they significantly improved classification accuracy, demonstrating the potential of AI to assist pathologists in this critical diagnostic task.

What Problem Does This Paper Address?

Ovarian cancer, a common and aggressive gynecologic malignancy, is subclassified into several subtypes (serous, mucinous, endometrioid, and clear cell). Accurately differentiating these subtypes from cytological images is crucial for proper diagnosis and treatment. However, this task can be challenging and time-consuming for pathologists, especially when dealing with a large volume of images, leading to potential errors. This paper aims to address this by developing an automated classification system using deep learning to aid pathologists.

Methods

The core of this research is the application of a Deep Convolutional Neural Network (DCNN) architecture, specifically inspired by AlexNet. The DCNN comprises five convolutional layers, three max-pooling layers, and two fully connected layers, with Rectified Linear Unit (ReLU) activation functions. To combat the common issue of insufficient training data in deep learning, the researchers employed data augmentation techniques. This involved two main strategies:

  • Image Enhancement: Applying a Gaussian High Pass-filter and a Laplass filter to improve image clarity and edge sharpness.
  • Image Rotation: Rotating the original images by 0°, 90°, 180°, and 270° to increase the sample size.

The model was trained and evaluated using two distinct datasets: one with original images and another with augmented images. A 10-fold cross-validation strategy was used to assess the classification accuracy.

Datasets

The study utilized a dataset of 85 ovarian cancer specimens, comprising 24 serous carcinoma, 22 mucinous carcinoma, 21 endometrioid, and 18 clear cell carcinoma cases. These were H&E stained tissue sections. Initially, 1848 original cytological images of uniform size (13601024 pixels) were obtained. These were then preprocessed into 227227 pixel images, resulting in 7392 original images. Through data augmentation (enhancement and rotation), the dataset was expanded to 81312 images, an 11-fold increase.

Key Insights and Ideas

  • Data Augmentation is Crucial: The study strongly emphasizes that augmenting the training data significantly boosts DCNN performance in medical image classification. Simply increasing the volume of data through enhancement and rotation had a tangible impact.
  • DCNNs Can Learn Complex Patterns: The DCNN, without explicit feature extraction or object segmentation, was able to learn to differentiate between ovarian cancer subtypes directly from pixel data. This highlights the power of deep learning in discovering subtle morphological differences.
  • Morphological Ambiguity is a Challenge: Misclassified images often shared common characteristics like blurred cell membranes/nuclei, overlapping cells, or mixed cell types within a single image. This suggests that the inherent visual ambiguity in some cytological samples remains a hurdle for even advanced AI models.

Key Results

The research found that the DCNN model trained with augmented image data achieved a classification accuracy of 78.20%, a substantial improvement over the 72.76% accuracy achieved by the model trained with original images. This represents an increase of approximately 5.44%. The study also noted specific misclassification patterns, such as endometrioid being mistaken for serous carcinoma (15.11% error rate), mucinous for clear cell (12.64%), and clear cell for mucinous (11.39%).

Limitations and Future Work

The authors acknowledge that some ovarian cancer images are misclassified, particularly those with unclear morphological features. They propose future work to address this by:

  • Increasing Sample Volume: Specifically focusing on augmenting samples with poor cell morphology.
  • Adjusting DCNN Architecture: Experimenting with different numbers of convolutional layers, filters, and max-pooling sizes.
  • Exploring Other Networks: Investigating architectures like GoogleNet, VGGS, and others to potentially improve classification accuracy.

The study also mentions that further immunohistochemical tests and manual readings by experienced pathologists are still necessary for misclassified specimens to ensure correct diagnosis.

Summary updated

Abstract

Ovarian cancer is one of the most common gynecologic malignancies. Accurate classification of ovarian cancer types (serous carcinoma, mucous carcinoma, endometrioid carcinoma, transparent cell carcinoma) is an essential part in the different diagnosis. Computer-aided diagnosis (CADx) can provide useful advice for pathologists to determine the diagnosis correctly. In our study, we employed a Deep Convolutional Neural Networks (DCNN) based on AlexNet to automatically classify the different types of ovarian cancers from cytological images. The DCNN consists of five convolutional layers, three max pooling layers, and two full reconnect layers. Then we trained the model by two group input data separately, one was original image data and the other one was augmented image data including image enhancement and image rotation. The testing results are obtained by the method of 10-fold cross-validation, showing that the accuracy of classification models has been improved from 72.76 to 78.20% by using augmented images as training data. The developed scheme was useful for classifying ovarian cancers from cytological images.
BibTeX
@inproceedings{wu2018automatic,
  author = {Miao Wu and Chuanbo Yan and Huiqiang Liu and Qian Liu},
  title = {{Automatic classification of ovarian cancer types from cytological images using deep convolutional neural networks}},
  booktitle = {Bioscience},
  year = {2018},
  doi = {10.1042/bsr20180289},
}
Figure