New machine learning model reduces the risk of detecting breast cancer

Breast cancer is the most prevalent cancer with the highest mortality rate. The speed of diagnosis and rapid detection decreases the severity of the disease. It is difficult to categorize breast cancer using histopathology photos that examine cells and tissues under a microscope – due to the nature of the data and inability to access large amounts of annotated data. Convolutional neural network (CNN) is a machine-learning technique that can detect breast cancer. However, it has shown promise.

Without any assurance these false predictions from CNN could lead to catastrophic outcomes. Researchers at Michigan Technological University have developed a machine-learning model that assesses the degree of uncertainty of CNN’s predictions. It classifies benign and malignant tumors, which reduces the risk.

In a paper that was recently published in the journal IEEE Transactions on Medical Imaging Mechanical engineering graduate students Ponkrshnan Thiagarajan and Pushkar Kharinar and Susanta Ghosh as assistant professor of mechanical engineering and machine-learning expert, describe their innovative probabilistic machine learning model, that is superior to similar models.

Machine learning algorithms that have been developed to date will still have uncertainty in their predictions. There is no way to quantify these uncertainties. Even if an algorithm informs us that a person is suffering from cancer, we do not know the degree of confidence in that prediction.”

Ponkrshnan Thiagarajan

Experience is the best way to gain confidence

In the medical context the uncertainty of knowing how reliable an algorithm is has made it difficult to rely on computer-generated predictions. The model described here is an extension of the Bayesian neural network – machine learning model that is able to evaluate an image and generate an output. The parameters of this model are considered to be random variables that allow for the quantification of uncertainty.

Michigan Tech’s model differentiates between positive and negative classes by analysing images. Images are at their simplest, collections of pixels. In addition to this classification, the model can determine the degree of uncertainty in its predictions.

In a medical laboratory the model could provide time savings by classifying images faster than an experienced lab tech. And, because the model can assess its own level of certainty and make recommendations on images to a human expert when it is not confident.

What is the reason an engineer from mechanical engineering creating algorithms for medical professionals? Thiagarajan’s idea was sparked when he first began making use of machine learning to cut the time required for solving mechanical engineering problems. Whether a computation examines the deformation of building materials or determines if someone is suffering from breast cancer, it’s crucial to know the uncertainties of the computation -; the key ideas remain the same.

“Breast cancer is one of the cancers with the highest mortality and incidence,” Thiagarajan said. “We believe that this is an exciting problem where better algorithms can have an an impact on people’s lives directly.”

Next steps

After their research has been published The researchers will now expand the model to include multiclass classification of breast cancer. Their aim will be to identify cancer subtypes, in addition to classifying benign and malignant tissues. The model is able to be expanded to include other medical diagnoses, even though it was developed from breast cancer histopathology images.

Ghosh stated that machine learning-based classification models can offer a lot but their predictions are not without uncertainty because of the inherent randomness of data and biases. “Our research attempts to address these issues and quantifying and analyzes the uncertainty.

Ultimately, Thiagarajan, Khairnar and Ghosh’s own model – that can determine whether images have low or high measures of uncertainty and identify instances where images require the attention of a medical professional; represents the next steps in the field of machine learning.

Journal reference:

Thiagarajan (P.), and. (2021). Explanation of Uncertainty derived from Bayesian Neural Network Classesifiers for Breast Histopathology Photographs. IEEE Transactions on Medical Imaging.

Content Source:

Gemma Wilson

Gemma is a journalism graduate with keen interest in covering business news – specifically startups. She has as a keen eye for technologies and has predicted quite a few successful startups over the last couple of years.

Related Articles