First-of-its-kind explainable AI model detects brain cancer

Source: Artvizual/Pixabay

One area of great hope and promise for deep learning applied to artificial intelligence (AI) lies at the intersection of neuroscience and oncology, two challenging areas known for their inherent complexity. A new study published in Biology methods and protocols demonstrates how the unique combination of Explainable AI (XAI) and repurposed camouflage animal detection algorithm can identify human brain cancer.

“This investigation is the first of its kind to apply camouflage animal transfer learning to training deep neural networks on a tumor detection and classification task,” wrote lead author Arash Yazdanbakhsh, MD, Ph.D., research assistant professor in the Department of Psychological and Brain Sciences at Boston University, in collaboration with Faris Rustom, Ezekiel Moroze, Pedram Parva and Haluk Ogmen.

Worldwide, cancers of the brain and central nervous system accounted for more than 321,000 new cases and 248,500 deaths in 2022, according to a report from the Global Cancer Observatory of the International Agency for Research on Cancer. world health.

About 90,000 brain tumors are diagnosed each year in the United States, of which about 25,200 are cancerous according to the Cancer Facts and Figures 2024 from the American Cancer Society; National Brain Tumor Society; Global Coalition for Adaptive Research. In 2024, there will be approximately 25,400 new cases of brain and other nervous system cancers and 18,760 deaths in the United States, according to the same report.

In the United States, an estimated one million people are living with a primary brain tumor, and approximately 28% of all brain tumors are malignant, according to the National Brain Tumor Society (NBTS). The deadliest form of brain cancer, glioblastoma (GBM), accounts for 50 percent of all primary cancerous brain tumors in America and has a median survival of eight months and a five-year relative survival rate of 6.9 percent by NBTS.

A brain tumor is the growth of abnormal cells in brain tissue. A primary brain tumor is a tumor that originates in the brain or brain region, as opposed to a metastatic brain tumor, which is cancer that has spread to the brain from another part of the body.

Primary brain tumors can be subdivided into glial tumors or gliomas and non-glial tumors. The human nervous system contains neurons and non-neuronal cells called glials. Neurons, also called nerve cells or neurons, are excitable cells that transmit electrochemical impulses. The human brain is made up of about 86 billion neurons, according to Suzana Herculano-Houzel, Ph.D., a neuroscientist at Vanderbilt University, in her 2012 article published in the Proceedings of the National Academy of Sciences of the United States of America (PNAS).

The word glia comes from the ancient Greek γλία meaning “glue”. Within the central nervous system (CNS), the major glial cell types include astrocytes, microglia, and oligodendrocytes; In the peripheral nervous system (PNS), satellite glial cells, enteric glial cells, and Schwann cells are examples of glial cells.

Glial cells are abundant in the human brain. Neuroscientist Nicola Allen of the Salk Institute for Biological Studies in La Jolla, California, and neurobiology professor David A. Lyons of the University of Edinburgh in Scotland estimate that glial cells make up about 50 percent of the central nervous system and play a role in the formation and function of the central nervous system. A previous estimate suggests that glial cells make up 90 percent of brain cells, according to a 2007 paper published in The International Journal of Biochemistry and Cell Biology by University of California, Los Angeles, researchers Fei He and Yi E. Sun.

Not all brain tumors are fatal. There are more than 150 distinct types of brain tumors identified, according to the American Association of Neurological Surgeons (AANS). According to the American Brain Tumor Association, about 27.9 percent, or about one-third, of all brain and central nervous system tumors in the United States are cancerous tumors, also called malignant tumors.

What sets this current study apart is the incorporation of an AI transfer learning stage, from detecting camouflaged animals to finding brain tumors in MRI images, by putting the emphasis on explainability and transparency.

“Although the tasks of detecting camouflage animals and classifying brain tumors involve different images, there could be a parallel between an animal hidden by natural camouflage and a group of cancer cells blending into surrounding healthy tissue “, the scientists wrote.

The researchers hypothesized that an AI network trained to spot camouflaged animals could be effectively repurposed to detect brain tumors from image data obtained non-invasively from brain imaging scans. magnetic resonance (MRI). In radiology, the two main types of MRI images are T1 and T2 weighted. T1-weighted images highlight fat and are ideal for normal soft tissue anatomy, while T2-weighted images are ideal for detecting fluid and abnormalities such as tumors, trauma and inflammation, according to the manual Merck 2023.

The scientists reused a pre-trained AI convolutional neural network (CNN) to identify camouflaged animals into two AI models, one for classifying T1-weighted MRI scans, called T1Net, and the other, called T2Net , to classify T2-weighted MRIs.

The team deployed the explainable AI techniques of a feature visualization method called DeepDreamImage, image saliency mapping, and feature spaces.

Brain scan data for gliomas came primarily from the NIH National Cancer Institute’s Cancer Imaging Archive (TCIA) and Kaggle public databases. Tumor categories included oligodendrogliomas, oligoastrocytomas, and astrocytomas. Data from normal MRI scans from de-identified patient records at the Boston VA Healthcare System were also used to train the artificial neural networks as a control.

“Both T1Net and T2Net had near-perfect accuracy on normal brain images, with only 1-2 false negatives between the two networks, demonstrating a strong ability to differentiate cancerous brains from normal brains,” the scientists reported.

Transfer learning from animal camouflage detection has strengthened AI’s ability to classify brain tumors, particularly astrocytomas. The transfer learning-enhanced AI model achieved 92.2% accuracy for the T2-weighted MRI model, which outperformed the model without transfer learning.

“Our results demonstrate that this approach to training deep neural networks is promising, particularly when using T2-weighted MRI data, as this model showed the greatest improvement in testing accuracy,” shared the researchers.

The researchers also report results that show that the qualitative XAI methods used enabled the visualization of what happens during AI training using brain cancer MRI data, as well as traits associated with various types of tumors. Through explainable AI methods, the decision-making process of glioma classification was shown to take into account a methodology based on tumor-specific features by the AI models.

AI’s deep learning renaissance is enabling significant scientific advances. Using AI to non-invasively distinguish between cancerous and non-cancerous brain tumors could provide an assistive tool for clinicians, oncologists and radiologists in the future.