A tumor is a collection of cells that grow faster than usual. This abnormality is either because of a problem in the death cycle or the division cycle of cells. The Brain tumor is a tumor that originates either within the brain or spreads itself from some other organs to the brain. Hundreds of thousands of patients belonging to different genders, races, and age groups each year are diagnosed with this fatal disease.

Like any other disease, the timely diagnosis of a brain tumor is critical for the doctors to treat the patient. The diagnosis may be made using invasive (biopsy) or non-invasive (examination of medical images, physical examination) methods. The biopsy is an invasive method and, therefore, expensive, time-taking, and risky, but it is also considered a gold standard. Examination of medical images of patients, like magnetic resonance imaging (MRI), is a non-invasive method and is relatively fast and inexpensive. MRI is the most suitable imaging modality for brain tumor diagnosis because it is not only the best modality for taking images of soft tissues but also the least harmful in terms of radiation. At the same time, the manual examination of various imaging modalities for brain tumor diagnosis lacks the desired accuracy.

Manual examination of medical images for brain tumor diagnosis is time-consuming for radiologists as they must process a lot of images/volumes every day. Moreover, it suffers from inter and intraobserver variability and hence lacks the desired accuracy. Machine learning-based computer-aided systems have been proposed to overcome this problem. These proposed systems perform the task of brain tumor type classification, brain tumor segmentation, and grading. In addition to diagnostic tasks, some also perform prognostic tasks like survival prediction of brain tumor patients.

With the advent of deep learning, most computer-aided systems now use deep learning models. A few popular architectures in this regard are convolutional neural networks (CNNs), U-Net models, fully convolutional neural networks (FCNNs), and sequential models. A recent addition in this regard is the transformer architecture. Although these deep learning-based systems can learn complex features from medical images and volumes, they require colossal data to avoid overfitting. Many researchers have used data augmentation techniques like flipping horizontally or vertically, rotating, and other similar methods. Researchers in the domain of medical imaging have also used semi-supervised learning to estimate labels for unlabeled data to have more labeled data for training. Similarly, self-supervised learning-based methods are also being proposed to use unlabeled data.

The current research proposes a complete framework for the diagnosis and prognosis of brain tumors. The tasks included in the diagnosis will be brain tumor type classification, brain tumor segmentation, and brain tumor grade classification, whereas the prognostic task will be the survival prediction for the patients. The study will propose state-of-the-art deep learning models to perform classification/segmentation/survival prediction by exploiting the maximum contextual information available in imaging modalities (and clinical data) while remaining within the computational constraints. Several techniques may be used to handle this data scarcity, including traditional data augmentation (like flipping and rotating of images/volumes), creating synthetic data using generative models like generative adversarial networks (GANs), increasing the number of labeled cases by using semi-supervised methods and/or using self-supervised methods for using unlabeled data (as it is costly to label medical images). The study will use publicly available datasets (like BraTS, FeTS, Whole Brain Atlas - Harvard Medical School (HMS)) as well as a private dataset like Bahawal Victoria Hospital, Bahawalpur, Pakistan (BVHB) dataset.  

Research Grants Won

We won the "NVidia Academic Hardware Grant 2022" worth 2400 US Dollars for the project "Computationally Efficient Domain Adaptive Pre-Training" which is a sub-part of our framework.

Drafts

Resource-efficient domain adaptive pre-training for medical images 

Yasar Mehmood, Usama Ijaz Bajwa, Xianfang Sun (arXiv)

Team Members