Tversky Loss Mechanisms: A ResUNet Approach to Improving Brain Tumor Segmentation

Biomedical imaging exemplifies a modern transdisciplinary collaboration in medicine. Various deep learning and AI techniques can automate tumor segmentation. Brain tumors typically have the poorest prognosis, as most patients succumb once the tumors reach their maximum size. It is crucial to identify and segment the tumor when it is small enough for surgical removal, despite the availability of multiple imaging modalities for tumor segmentation. Magnetic Resonance Imaging (MRI) produces a multidimensional image of the brain that requires segmentation. However, this is a labor-intensive manual task that often exhibits significant inter-observer variability. Automated segmentation will therefore enable faster and more efficient tumor segmentation for accurate early identification of brain tumors. This study introduces the ResUNET segmentation network utilizing a Tversky loss function. It tackles class imbalance, a significant challenge in brain tumor segmentation. We surpass UNET in segmentation outcomes by addressing class imbalance and accurately segmenting the smaller, critical areas of the tumor. Our model achieves an intersection-over-union (IoU) score of 0.95 and a predicted Dice coefficient of 0.98 during testing, indicating a highly accurate match for expected tumor sites with a deviation of only 0.1 in placement. In other words, the tumor segmentation maps we predict reflect the size of the input images, making them both valid and significant. These results illustrate the progress our method has made in automating brain tumor segmentation, aiding in early diagnosis and treatment.