New Segmentation Tool Enables Healthcare Professionals to Teach Computers How to Correctly Annotate Medical Images

The images are worth a thousand words, but with medical images, it's a euphemism. Digital images of biopsies are particularly useful for diagnosing and tracking the progression of certain diseases, such as chronic kidney disease and cancer.

Computer tools called neural networks, which focus on the recognition of complex shapes, are well suited to these applications. But because machine learning is so complex, health professionals typically rely on computer engineers to "train" or modify neural networks to properly annotate or interpret medical images.

Researchers at the University of Buffalo have developed a tool for healthcare professionals to badyze images without technical expertise. The tool and image data that have been used for its development are publicly available at: https://github.com/SarderLab/H-AI-L.

The technique has been described in an article published in Nature Machine Intelligence On February 11, researchers planned a tool to digitize medical images of any medical organ, with histological images of chronic nephropathy and magnetic resonance images of the human prostate.

"We have created an automated human segmentation tool in the loop for pathologists and radiologists," said Pinaki Sarder, Ph.D., corresponding and principal author, and badistant professor in the Department of Pathology and Anatomic Sciences at the School Center. of Medicine and Biomedical Sciences at UB. The lead author of the paper is Brendon Lutnick, Ph.D. candidate at the Jacobs School and in charge of his doctoral thesis under the supervision of Sarder.

Intuitive interface

Designed with what researchers call an intuitive interface, the tool automatically enhances the annotation and segmentation of medical images based on what it "learns" from the way the user human interacts with the system.

"With our system, you do not have to know machine learning," said Sarder. "Now health professionals can annotate the structure themselves.

"This technique allows healthcare professionals to use for the first time their own tools, such as a complete slide display commonly used for annotating images, without getting lost in translation of machine learning lingo, "he said.

Lutnick explained that the system is designed to improve its performance because it is "trained" on the same dataset. "You want to train it iteratively on your own data set," he explained. "This optimizes the workload of the expert annotator as the system becomes more efficient with each use."

The system improves iteratively, learning primarily whenever the health professional redraws a border on an image to identify a particular structure or anomaly.

A better way to predict the progression of the disease

The ultimate goal is a more accurate understanding of the patient's pathological condition. "When you do a biopsy, you want to understand the features of the image and what they tell you about the progression of the disease," said Sarder.

He explained that, for example, a darker red area on an image of the glomerulus in the kidney, where the waste is filtered from the blood, indicates sclerosis, which may indicate that the disease has progressed. The more precisely the boundaries of these areas can be defined, the better is the understanding of the stage of the disease in which the patient is and how it can evolve in the future.

"The system works better every time," said Lutnick, "so the burden of the man who uses the machine is reduced at each iteration." Every time the individual redraws a border on a sample, the This learning system allows humans to understand the weaknesses of the machine as it learns. "

Provided by
University in Buffalo