Machine learning to aid in classification of pathological images for disease diagnosis

The Dual-Channel Prototype Network (DCPN) aims to classify pathological images efficiently based on limited available data to effectively categorize (and even diagnose) rare diseases

Identification and intervention when it comes to disease diagnosis can be the result of someone’s observations through a microscope. However, microscopy can be time-consuming and labor-intensive when pathologists have to observe tissue samples to make a diagnosis, not to mention the chance for human error or bias. In the case of a rare disease, the issue is further complicated. This is why intervention is needed to develop high-quality datasets based on real-world clinical settings. With the use of few-shot learning (FSL) to make the best out of a limited data situation, machine learning can help analyze and classify patient data to create a preferred method of establishing categories for patient diagnosis through pathological images, especially when it comes to rare diseases.

Researchers published their results in IEEE Journal of Biomedical and Health Informatics on June 7.

“The Dual Channel Prototype Network (DCPN) utilizes a self-supervised learning strategy to

pre-train the transformer model and, when combined with convolutional neural networks,

can effectively extract multi-scale prototype representations of images, significantly enhancing the generalization ability of the prototype representations,” said Xiaoyu Cui, author and researcher on the study.

The study introduces a three-part method using the machine-learning technique FSL to help with the classification of pathology images. FSL is a machine-learning technique that trains models to make predictions, most notably when data is scarce. The first part consists of pretraining the pyramid vision transformer (PVT) network for the purpose of medical image classification. PVT is an application often used for the detection and classification of images. A PVT can take small components of data and reduce the length of the information to reduce computational cost. Then, the DCPN is constructed by using a convolutional neural network (CNN). A CNN is a commonly used machine learning program that attempts to mimic the human brain in decision-making (artificial neural network) and is best used for recognizing patterns in images. The third part uses algorithms designed to combine the probability of predictions of each classifier (which is essentially a categorizing algorithm) known as “soft voting” and multi-scale features (information on a feature that is captured at different spatial scales within an image) to fulfill the parameters of few-shot classification. This stepwise process completes the DCPN method.

The experimental results, run on three public datasets, indicate a noticeable advantage when using DCPN compared to other methods. Additionally, when it came to tasks within the same domain (for example, data from the same data source and organ), DCPN ranks very similarly to traditional supervised learning methods. Although the prototype decreases in performance when more complex cross-domain tasks are involved, this can be improved upon with later training enhancements and adaptations.

In the future, the DCPN method will ideally be employed in a large-scale manner to bring an effective few-shot pathological image classification algorithm into use widely across different clinical settings. While there is some work to be done to improve the performance and robustness of the model, the hope is with improved imaging and classification, the meager amounts of annotated pathological images will no longer be a major hurdle in the face of medical progress and the diagnosis of less common diseases.

Hao Quan, Xinjia Li, Tianhang Nan and Xiaoyu Cui of the College of Medicine and Biological Information Engineering at Northeastern University and Dayu Hu of the School of Computer at the National University of Defense Technology contributed to this research.

The China Key Research and Development Program, the Natural Ningbo Science and Technology Bureau, the Fundamental Research Funds for the Central Universities and the Liaoning Province Medical Engineering Cross Joint Fund made this research possible.