Automatic Brain Tumour Detection and Segmentation using Machine Learning

A leading international hospital

The Challenges

Brain tumors occur because of abnormal growths and uncontrolled cell divisions in the brain. The determination of the tumor extract poses a major challenge in brain tumor treatment planning and quantitative evaluation. Manual segmentation of brain tumor extract from 3D MRI volume is time consuming while its performance is highly depending on experience of the operator. In this context, the client approached Guires to perform a reliable fully automatic segmentation method for the brain tumor segmentation. Automatic segmentation algorithms that can achieve results that are used to support doctors more easily in diagnosing and treating the patients for detection and segmentation of the tumour.

Our Strategy

Our team of data science experts collected MRI mages from hospitals. In collaboration with physicians and medical experts’ team, a protocol was developed that define the objective of this study, research questions, hypotheses, methods and techniques proposed for the analyses, sampling, sample size, and endpoints expected. The first objective was to deals with the detection of the tumor from MRI images while the second part contains the process of classification of tumor type (Benign, Malignant or Normal). The input images had undergone a number of stages including pre-processing, segmentation and classification.

Pre-processing using normalization and histogram matching was carried out. Features were extracted to identify the texture and color feature. Several data mining algorithms such as SVM, K-means, k-nearest neighbour, FCM, a neural network have been used to identify the best method.  Data augmentation techniques (such as flip horizontally, flip vertically, rotation, shift, shear, zoom, brightness and elastic distortions) were applied to improve the network performance by intentionally producing more training data from the original one. The output was evaluated through PSNR, Mean and Standard Deviation, Structural Similarity Index Metric for the segmented images to test the percentage level of the accuracy.

Our Outcomes and Impact

We applied relevant CNN to classify the MRI images. The Fuzzy C means clustering, PCA method was used to classify the MRI Images. The output produced was the whole volume and provides a way to model tissue classes, which is more robust scheme under noisy or bad intensity normalization conditions which produce better results using high-resolution images, outperforming the results provided by other algorithms in the state-of-the-art, in terms of the average overlap metric.