Biomarkers identified from medical images are valuable for disease diagnosis and staging. Object detection and segmentation are important for improving the accuracy of biomarker identification. It can be challenging to detect objects, especially small objects or 'blobs,' from 3D images secondary to low image resolution, noise, and overlap. Traditional small blob filters generate significant false positives as noise components can be misidentified as regions of interest. Recent advancements in U-Net-a deep learning model-have yielded more effective denoising capabilities that make it more suitable for small blob detection. However, unless the intensity threshold is appropriately established, U-Net tends to 'under-segment To address under-segmentation, we propose a combination of U-Net paired with optimal threshold detection via Otsu's thresholding. Two sets of experiments have been performed to compare this approach with both Hessian-based blob detection and U-Net with standard thresholding. The first experiment evaluated 20 simulated 3D images-comprising gloms with a varied size distribution, and noise. The second experiment examined MR images from 11 mouse kidneys with the objective of detecting all glomeruli. Our results support the conclusion that the proposed U-Net with optimal thresholding outperforms the Hessian-based detection and U-Net with standard thresholding approaches.