This paper talks about a new, multi-step segmentation method for medical image analysis that aims to make radiological image segmentation more accurate and reliable. The suggested plan uses thresholding methods with advanced image processing methods like Gaussian smoothing, region expanding, and class-specific thresholding refinement to make the quality of segmentation better. A clinically focused, multi-step radiological image segmentation pipeline is shown. It uses traditional thresholding, class-specific refinement, and modern image processing to make results that are strong and well-calibrated. The method uses normalization and histogram equalization to make the intensities the same. It then uses Gaussian smoothing to make an original mask that is changed over and over again with Otsu optimization, edge cues, region growth, and morphology. Benchmarks that compare different modalities and body parts to strong classical and deep baselines show the best overlap and boundary accuracy in a real-world runtime. The dice had a precision of 0.90, an IoU of 0.83, an HD95 of 4.6 ± 2.8 mm, an ASSD of 0.88 ± 0.38 mm, and a recall of 0.92. The reliability study shows better calibration (ECE 3.8%, Brier 0.098, NLL 0.58) and almost no volumetric bias in the Bland-Altman comparison, which means the system is ready to be integrated into clinical workflow.