The Mekong Delta is Vietnam's most crucial rice production region. Rice grain shape quality plays a pivotal role in establishing commercial standards. Grain morphology, characterized by length and width, is one of the key rice quality parameters. This characteristic could be influenced by grain position within the panicle. Numerous studies have employed image processing techniques and machine learning models for rice grain detection, counting, measurement, and classification. However, no prior research has evaluated grain size variation across rice panicles. This study utilizes two machine learning models, YOLOv8 and DeepLabv3, to detect and segment rice grains and brown rice grains (after dehulling) in three panicle sections: top, middle, and bottom. From grain segmentations, grain sizes (length and width) were calculated as rectangles around the convex hull which are the smallest convex boundaries. Experimental results on 11 rice varieties of varied grain lengths, planted in the Vietnam Mekong Delta region, indicate no significant size differences between the sections. YOLOv8 provides better grain detection results compared to DeepLabv3 for images with adjacent grains, while DeepLabv3 gives more accurate grain boundary segmentation results. In addition, we used ANOVA to compare the evaluation results acquired using the two deep learning models to those obtained using the ruler method.