Probabilistic hierarchical clustering based identification and segmentation of brain tumors in magnetic resonance imaging

The automatic segmentation of the abnormality region from the head MRI is a challenging task in the medical science domain. The abnormality in the form of the tumor comprises the uncontrolled growth of the cells. The automatic identification of the affected cells using computerized software systems is demanding in the past several years to provide a second opinion to radiologists. In this paper, a new clustering approach is introduced based on the machine learning aspect that clusters the tumor region from the input MRI using disjoint tree generation followed by tree merging. Further, the proposed algorithm is improved by introducing the theory of joint probabilities and nearest neighbors. Later, the proposed algorithm is automated to find the number of clusters required with its nearest neighbors to do semantic segmentation of the tumor cells. The proposed algorithm provides good semantic segmentation results having the DB index-0.11 and Dunn index-13.18 on the SMS dataset. While the experimentation with BRATS 2015 dataset yields Dice(complete)=80.5 %, Dice(core)=73.2 %, and Dice(enhanced)=62.8 %. The comparative analysis of the proposed approach with benchmark models and algorithms proves the model's significance and its applicability to do semantic segmentation of the tumor cells with the average increment in the accuracy of around +/- 2.5 % with machine learning algorithms.

Automated summary

This paper introduces a new clustering approach based on machine learning for automatic identification and semantic segmentation of tumor regions in head MRI images. Experimental results demonstrate improved accuracy compared to benchmark models and algorithms.