Photo by @jcgellidon

Bayesian Convolutional Neural Network

A Step Forward in Establishing Accurate and Reliable Machine Learning Models for Histopathology

I decided to check out an article about called: ARA: accurate, reliable and active histopathological image classification framework with Bayesian deep learning [1]. This is not a review, I am simply attempting to summarise some of the points and explore a few words.

What is histopathological?

“Histology is the study of tissues, and pathology is the study of disease. So taken together, histopathology literally means the study of tissues as relates to disease. A histopathology report describes the tissue that has been sent for examination and the features of what the cancer looks like under the microscope.” [2]

This pertains to images routinely generated in clinical practice. They propose a new Bayesian Convolutional Neural Network (ARA-CNN) for classifying histopathological images of colorectal cancer.

“Colorectal cancer (CRC), also known as bowel cancer and colon cancer, is the development of cancer from the colon or rectum (parts of the large intestine). A cancer is the abnormal growth of cells that have the ability to invade or spread to other parts of the body.”

Reliability in Machine Learning

They argue that the advent of digital pathology made new images available for automated analysis at scale. However while accuracy is optimised by every machine learning (ML) method, reliability is another desired feature that is not delivered by many state of the art solutions.

“The analysed dataset holds 5000 image patches belonging to eight balanced classes of histopathologically recognisable tissues”

Since 2012, when the groundbreaking AlexNet was created by Alex Krizevsky ‘state of the art’ turned algorithms using manual feature engineering to deep learning. The architecture of the model, called ARA-CNN, was inspired by many state of the art solutions, including Microsoft ResNet and DarkNet 19.

Overview of the proposed ARA framework. (A) Active histopathology workflow. Annotated whole-slide images (WSIs) are split into small image patches, which constitute a dataset. ARA-CNN is trained on that dataset. After the first round of training, the pathologist should be informed about i) which classes are the most uncertain and ii) which image patches are misclassified and highly certain, and thus identified as potentially mislabelled. The former should inform the pathologist about which classes to prioritise in the next round of annotation. The latter should inform about which image patches should be re-annotated with correct labels. We then take new annotated whole-slide images and continue the workflow until we reach a satisfying level of classification accuracy. (B) Segmentation workflow. Whole slide images are split into small image patches. Each of these is classified by trained ARA-CNN and is assigned a colour based on its classification result. These coloured tiles are merged together to form a segmented whole slide image and can be analysed in terms of their spatial relationships. Each resulting tile has a measured uncertainty value as well, so pathologists can make an informed decision whether to take the automated segmentation as-is or to inspect it manually.

As you can see the pathologist plays an important role here.

“Pathology is the study of the causes and effects of disease or injury. The word pathology also refers to the study of disease in general, incorporating a wide range of bioscience research fields and medical practices.”

Bayesian Convolutional Neural Network (ARA-CNN) was applied to the task of colorectal tissue classification and incorporated it into an uncertainty-based active pathology workflow.

• To achieve reliability, the model measures the uncertainty of each prediction
• Their analysis involved a comparison of two different uncertainty measures, Entropy H, and BALD.
• The two measures agreed on which classes are most uncertain on average, pointing to classes which were most often misclassified by the model.
• Using H, the classification accuracy equal to that of the model trained with the full dataset was reached 45% faster

According to the paper: “The excellent performance of ARA-CNN indicates that it is a step forward in establishing accurate and reliable machine learning models for histopathology […] As future work, we plan to apply our model to other histopathological tissue datasets. Due to its deep learning nature, our architecture should easily handle tissue types other than colorectal (potentially with the help of transfer learning).”

1. https://www.nature.com/articles/s41598-019-50587-1?fbclid=IwAR2wVpRRDSIlOavDEFjdaBqNU_T1xVhmJiL736ZubIb0H0m_gZAZdSwmT7c
2. https://www.verywellhealth.com/histopathology-2252152

This is day 124 of #500daysofAI. If you enjoy this article please give me a response as I do want to improve my writing or discover new research, companies and projects.