Abstract
Brain tumor segmentation is a fundamental task in medical image
analysis because it supports clinical diagnosis, treatment planning, and
monitoring of disease progression. However, accurately delineating tumor
regions from Magnetic Resonance Imaging (MRI) scans remains challenging
due to the heterogeneous intensity characteristics of brain tissues and the
reliance of many segmentation approaches on large annotated datasets and
high computational resources.
This study proposes an unsupervised segmentation framework called
Adaptive Clustering with Morphological Refinement (ACMR) for automatic
brain tumor segmentation using Fluid Attenuated Inversion Recovery (FLAIR)
MRI images. The primary objective of the proposed framework is to reduce
the dependence on manual annotation while maintaining reliable
segmentation performance and low computational requirements. The ACMR
framework integrates several complementary components. First, Fuzzy CMeans (FCM)-guided Contrast Limited Adaptive Histogram Equalization
(CLAHE) performs adaptive contrast enhancement based on tissue intensity
distributions, improving the visibility of tumor boundaries. Second, a multiconfiguration clustering strategy is introduced to address the inherent
intensity variability in brain MRI images, which typically contain multiple tissue
types including tumor, gray matter, white matter, Cerebrospinal Fluid (CSF),
and background. Third, an adaptive entropy-weighted fusion mechanism
based on Shannon entropy combines segmentation outputs from different
clustering configurations to improve segmentation robustness. Finally,
morphological refinement is applied to enhance the anatomical consistency
of the final segmentation results.
X
The proposed framework was evaluated independently on the BraTS
2017 and BraTS 2021 benchmark datasets. Experimental results demonstrate
that the ACMR framework achieved a Dice Similarity Coefficient (DSC) of
81.02% on BraTS 2017 and 81.06% on BraTS 2021, while processing each MRI
slice in approximately 0.2 seconds. These results demonstrate that
unsupervised tumor segmentation with competitive accuracy can be achieved
without relying on manually annotated training data.
Overall, the proposed ACMR framework provides an effective and
computationally efficient unsupervised solution for brain tumor
segmentation, making it suitable for medical imaging environments where
annotated datasets and high-performance computing resources are limited.
