Biopsies are always the first call when it comes to diagnosing a case of brain cancer. Surgeons start by removing a thin layer of tissue from the tumor to find signs of disease closely under a microscope. Although biopsies are very presumptuous, the samples collected only represent a snatch of the whole tumor. MRI is a less bold but time-consuming process as radiologists have to manually map out the tumor area from the scan before the classification.
A team of researchers from Washington University have proposed a convolutional neural network (CNN) model to classify various types of tumors without biopsies. A convolutional neural network (CNN) is the model that uses deep learning, which is a type of machine learning algorithm available in image recognition software. Based on hierarchical attributes such as location and morphology, this model is used to recognize the tumors in MRI. The CNN model can classify various brain cancers precisely without any manual interplay.
According to the author, this is the first step toward developing artificial intelligence, which elevates radiology workflow that can aid image interpretation by furnishing quantitative information statistics.
The model created was able to categorize postcontrast T1-weighted MRI scans of several brain tumor types and distinguish pathologic from healthy scans. The CNN can determine six common types of intracranial tumors: high- and low-grade gliomas, meningioma, pituitary adenoma, acoustic neuroma, and brain metastases. As per the team of researchers, this neural network is the first to determine tumor class and detect the absence of a tumor from a 3D magnetic resonance volume.
The researcher has created two multi-institutional datasets of pre-operative and post-contrast MRI scans from four publicly available datasets besides data obtained at Washington University School of Medicine (WUSM) to find out the accuracy of the CNN.
Firstly, the internal datasets contain 1757 scans covering seven imaging classes, out of which six are tumor classes and one healthy class. Out of these scans, 1396 scans were used by the team as training data to train CNN to distinguish between each type. The remaining 361 scans were eventually used to test the model’s performance, which was the internal data.
As confirmed by the radiology reports associated with each scan, the CNN correctly identified tumor type with 93.35% accuracy. Moreover, the odds that a patient actually had specific cancer that the CNN detected was 85% – 100% rather than being healthy or having any other type of tumor. However, a small number of false negatives were found covering all imaging classes. The odds were that patients who tested negative for a given category were not having that disease or were not healthy was 98% – 100%.
After this, the researchers tested their model with a second external dataset containing only high and low-grade gliomas. These scans were not included in the internal dataset because they were sourced separately.
As per the researcher, deep-learning models are so sensitive to input. It’s become standard practice to test their performance on a separate dataset gathered from a different source to determine how well they generalize and react to new data.
On the external test data, the CNN showed good generalization abilities, scoring 91.95 percent accuracy. The findings show that the model may help doctors diagnose patients with the six tumor types investigated. The researchers admit that their approach has several flaws, such as misclassification of tumor kind and grade due to weak picture contrast. Inconsistencies in the imaging procedures utilized at the five institutions could be responsible for the deficiencies.
In the future, the team wants to improve CNN’s training by adding more tumor types and imaging modalities.