Abstract:
Gliomas are a type of brain tumor that affect the glial cells in the brain. Based on the severity, gliomas are classified as either low grade and high grade grade glioma. The treatment administered to a subject is dependent on various factors such as grade of the glioma, extent of tumor spread, location of the tumor, age of the patient etc. Segmentation of gliomas from the medical images provides insights about the extent of the lesion and hence is considered as the primary step in a clinical setting. Manual segmentation of lesions from medical volumes is cumbersome and introduces inter-rater variability. For good generalization on unseen data, machine learning based techniques need to be trained on a vast and diverse dataset. Access to large volumes of annotated data in the field of medical image analysis is rare. The thesis explores two automated techniques for the task of segmentation of gliomas from medical volumes. The techniques differ based on the requirement of labelled data during the training phase.
A semi-supervised technique pre-trained on large volumes of unlabeled data is finetuned on an extremely small cohort of labelled data for segmenting gliomas. As a measure to reduce the false positives generated by this segmentation network, another model trained in an unsupervised fashion is utilised to provide spatial priors about the presence of lesions in a volume. The second approach takes advantage of the amount of labelled data when it comes to segmentation of glioma. A variety of 2-D and 3-D based convolutional neural networks were developed in a supervised fashion for the task of segmentation of gliomas.
Using the segmentation generated by the networks, the thesis also explores making use of various machine learning algorithms to perform downstream applications such as estimating type of tumor, grade of the lesion, and also estimating the prognosis of a subject. An automatic and non-invasive technique was developed to differentiate high grade gliomas from low grade gliomas directly from MR volumes. The proposed pipeline was observed to achieve a performance comparable to trained radiologists.
From subjects diagnosed with glioblastoma multiforme, shape and texture based features from various lesion constituents were input to a gradient boosting algorithm for survival rate prediction. The features used frequently by the proposed algorithm were observed to be similar to the ones used by doctors. At present, the task of survival prediction is a problem statement associated with limited annotated samples. The thesis compares the performance of a standard machine learning approach which makes use of hand coded features against deep learning based approaches which generates features by learning from the data.
Finally, the thesis also introduces a Recurrent Attention Mechanism based algorithm to differentiate between various types of brain tumor from Magnetic Resonance Images. When compared to conventional deep learning techniques such as Convolutional Neural Networks, the proposed pipeline is more interpretable and computationally efficient. The lowered computation requirement ensures wider adoption of AI based solutions in hospitals wherein accessibility to high end compute facilities are rare.
Publications:
[1] Alex V., Vaidhya K., Thirunavukkarasu S., Kesavadas C., & Krishnamurthi G., "Semisupervised learning using denoising autoencoders for brain lesion detection and segmentation”, Journal of Medical Imaging, Vol.4, pp 041311, 2017
[2] Vaidhya K., Thirunavukkarasu S., Alex V & Krishnamurthi G., "Multi-modal brain tumor segmentation using stacked denoising autoencoders”, International Workshop on Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries, Munich, 2015.
[3] Alex V. ,Safwan M. & Krishnamurthi G., Automatic Segmentation and Overall Survival Prediction in Gliomas Using Fully Convolutional Neural Network and Texture Analysis, International MICCAI Brainlesion Workshop, Quebec, 2017.
[4] Shaikh M., Anand G., Acharya, G., and Amrutkar, A., Alex, V. & and Krishnamurthi G., Brain Tumor Segmentation Using Dense Fully Convolutional Neural Network, International MICCAI Brainlesion Workshop, Quebec, 2017.
[5] Alex V., Mohammed S., Chennamsetty S., & Krishnamurthi G. , Generative adversarial networks for brain lesion detection, Medical Imaging 2017: Image Processing, Orlando,2017.
[6] Kori A, Soni M, Pranjal B, Khened M, Alex V, Krishnamurthi G. Ensemble of fully convolutional neural network for brain tumor segmentation from magnetic resonance image. International MICCAI Brainlesion Workshop, Granada, 2018.
[7] Shaikh M, Kollerathu VA, Krishnamurthi G., Recurrent Attention Mechanism Networks for Enhanced Classification of Biomedical Images, IEEE 16th International Symposium on Biomedical Imaging (ISBI) Venice, 2019.