Investigation of the Medical Image Modalities to Detect the Abnormalities in the Spinal Cord Using Thresholding Techniques

Accurately segmenting spinal structures from magnetic resonance imaging (MRI) is essential for diagnosing degenerative disc diseases. However, 1.5 T low-field MRI often suffers from high noise and low contrast, complicating the segmentation process. This study provides a systematic comparative evaluation of four thresholding-based techniques: binary/global, Otsu, adaptive, and standard deviation-based methods. Using a dataset of spinal DICOM images, we analyzed the effectiveness of these algorithms in delineating bone and soft tissue boundaries. The performance was quantitatively validated using the Dice coefficient and statistical analysis via ANOVA. Results indicate that adaptive thresholding and Otsu thresholding outperformed global methods in handling localized intensity variations, achieving a higher similarity index to expert-validated reference masks. Statistical analysis using ANOVA (p > 0.05) further suggests that while these traditional methods are computationally efficient, their efficacy is constrained by the signal-to-noise ratio (SNR) in low-field imaging. This research establishes a technical benchmark for traditional segmentation limits, highlighting the specific scenarios where a transition to deep-learning-based models becomes necessary for clinical accuracy.