Multi-classification of Brain Tumor MRI images using Convoluted Neural Network (CNN) based approach

The human brain is a complex organ, and when abnormal tissue begins to grow within it, the consequences can be severe. Doctors rely heavily on magnetic resonance imaging, or MRI, to see inside the skull without surgery. These scans produce detailed pictures that reveal the shape and location of a tumor. However, reading these images is a demanding task. A single patient's scan can contain dozens of slices, and a radiologist must examine every one to spot subtle differences between healthy tissue and various types of cancer. This process takes time, and even the most skilled experts can miss a detail or misidentify a tumor type when faced with thousands of images. To help, scientists have turned to computers, teaching them to recognize patterns in these pictures just as a human eye would. This field, known as deep learning, uses artificial systems that learn by example rather than by following a fixed set of rules. The goal is to create a tool that can assist doctors in diagnosing brain tumors faster and more accurately, potentially saving lives by catching diseases earlier.

In a recent study, researchers from the University of Delhi tackled this challenge by building a specialized computer program designed to sort brain MRI scans into four distinct categories. They wanted to know if their system could distinguish between a healthy brain and three specific types of tumors: gliomas, meningiomas, and pituitary tumors. Gliomas are tumors that start in the glue-like tissue supporting the brain, meningiomas grow on the membranes covering the brain, and pituitary tumors form near the pituitary gland at the base of the skull. The team did not use a pre-existing, massive computer model that had already been trained on millions of images. Instead, they constructed a custom, lightweight system from scratch. This approach was chosen to keep the technology simple and efficient, requiring less computing power and memory than the heavy, complex models often used in other studies.

The researchers began by gathering a large collection of brain images from a public database. They selected 7,023 pictures in total, which included 1,621 gliomas, 1,645 meningiomas, 1,757 pituitary tumors, and 2,000 images of healthy brains with no tumors. These images came from different angles—some looking from the top down, others from the side or the front—providing the computer with a diverse view of the brain's anatomy. Before teaching the computer, the team prepared the images by resizing them all to the same dimensions and adjusting the brightness levels so the computer could process them more easily. They then split the data, using the majority of the images to teach the system and holding back a smaller group to test how well the system learned.

The computer model they built works by peeling away layers of information, much like looking at a picture through a series of increasingly specific lenses. The system contains seven layers that scan the image to find basic shapes and textures, followed by layers that combine these simple details into more complex patterns, such as the boundaries of a tumor. The researchers used a specific mathematical method to help the system learn quickly and avoid getting stuck, and they included a mechanism to prevent the system from memorizing the training images too perfectly, which would make it fail on new data. After training the model for 25 rounds, where it saw the images repeatedly and adjusted its internal settings to improve its guesses, the results were striking.

When the researchers tested the model on the images it had never seen before, it correctly identified the type of brain condition in 96.1 percent of the cases. The system was particularly good at spotting healthy brains and pituitary tumors, achieving near-perfect scores for those categories. It also performed strongly on the more difficult glioma and meningioma types, correctly classifying them the vast majority of the time. The study compared these results to other recent attempts to solve the same problem, many of which relied on much larger and more complex computer architectures. Those previous models achieved accuracy rates ranging from roughly 89 percent to 95 percent. The new, simpler model outperformed them all, reaching 96.1 percent accuracy while using fewer resources.

The researchers noted that the system occasionally confused gliomas with meningiomas, which is understandable because these two types of tumors can look very similar on an MRI scan. However, the overall performance was robust enough to suggest that this lightweight approach is a viable tool for medical use. The authors emphasize that while the computer model shows great promise, it is not yet a replacement for a doctor. The next steps involve testing the system on data from multiple hospitals to ensure it works reliably across different types of MRI machines and patient populations. If future validation confirms these findings, such a tool could become a standard part of the diagnostic process, helping medical teams make faster, more confident decisions about how to treat patients with brain tumors.