Multilevel Thresholding Fuzzy C-Means (MTFCM) hybrid method for osteosarcoma necrosis extraction in magnetic resonance imaging

Osteosarcoma (OS) is a malignant bone tumor primarily affecting adolescents and young adults. Tumor necrosis, indicative of dead tumor cells, plays a critical role in determining prognosis and guiding treatment strategies. MRI is a vital imaging tool for evaluating necrosis due to its high resolution and ability to distinguish soft tissue details. Manual interpretation of MRI images is labor-intensive, subjective, and timeconsuming, requiring radiologists to compare multiple sequences such as T1, T2, and T1 GADO. Intensity inhomogeneity refers to uneven signal intensity distribution in MRI images, which produces ambiguous boundaries and can adversely affect image interpretation during diagnosis. Digital image processing offers an alternative approach to assist manual interpretation. Traditional image processing algorithms, such as Fuzzy C-Means (FCM) and Multilevel Thresholding (MT), are widely used but come with specific limitations. FCM is effective in clustering regions based on intensity but struggles with overlapping intensities and computational demands, while MT offers efficient segmentation but lacks precision in handling complex tissue boundaries. In this study, two methods are proposed: the Adaptive Disk Structure Element Morphological (ADSEM) method for correcting intensity inhomogeneity and extracting fat regions, and the integrated Multilevel Thresholding Fuzzy C-Means (MTFCM) method for necrosis extraction using T1, T2, and T1 GADO sequences. The ADSEM method compensates for intensity variations in the T1 sequence and enables subsequent fat suppression in T2. The MTFCM method combines the strengths of MT and FCM to improve segmentation accuracy in regions with overlapping intensities. Total number of data includes 223 MRI slices. Validation was performed across multiple MRI slices, where the method was assessed consistently against radiologist-annotated ground truth. The ADSEM method achieved an accuracy of 89.73%, precision of 87.56%, recall of 70.76%, F1 score of 77.35%, and a Dice Similarity Coefficient (DSC) of 85.94% in fat extraction, while MTFCM delivered enhanced results in necrosis extraction with an accuracy of 99.03%, precision of 85.62%, recall of 62.30%, F1 score of 71.57%, and a Pearson correlation of 91.51% with the radiologists’ ground truth. In addition, boxplot analysis demonstrates a relatively narrow data range with concentrated distributions, indicating stable performance and consistent alignment with the radiologists’ ground truth across the evaluated MRI slices.