MozzaVID: Mozzarella Volumetric Image Dataset

1. Problem Statement

The field of volumetric deep learning faces a critical bottleneck: a severe shortage of large-scale, clean, and versatile benchmark datasets.

• Data Scarcity: Existing volumetric datasets (primarily in medical imaging) are typically small (100–10,000 volumes), often restricted by privacy or acquisition complexity, and usually tied to specific, complex scientific questions rather than general method development.
• Benchmarking Gap: Unlike 2D computer vision, which relies on massive, standardized datasets (e.g., ImageNet, COCO) to evaluate and compare architectures, volumetric research lacks a common ground. This forces researchers to adapt 2D models to specific 3D tasks without robust benchmarks, limiting the development of architectures optimized specifically for volumetric data.
• Food Science Gap: There is no large, publicly accessible 3D dataset dedicated to food microstructure, despite the importance of understanding complex, disordered food structures for developing sustainable alternatives (e.g., plant-based meats).

2. Methodology and Dataset Construction

Data Acquisition:

• Source: The dataset consists of high-resolution X-ray Computed Tomography (CT) scans of mozzarella cheese microstructure.
• Imaging Technology: Synchrotron radiation (DanMAX beamline, MAX IV) was used to overcome the low X-ray attenuation of proteins and fats, enabling fast, high-resolution (0.55 µm pixel size), low-noise scans.
• Samples: 25 distinct mozzarella cheese types were produced with varying recipes (cooking temperature, screw speed, additives). From these, 149 unique samples were prepared (6 samples per cheese type, except for one isotropic control).
• Scanning: Each sample underwent 4 local tomography scans, resulting in an initial pool of 600 scans. After discarding 9 scans with artifacts, 591 high-quality scans remained.

Dataset Preparation (The Three Instances):
To bridge the gap between small typical volumetric datasets and massive 2D datasets, the authors created three dataset instances by downsampling and splitting the raw volumes (originally ~1600×1600×2156 voxels) into standardized 192×192×192 voxels:

1. Small (8X-1X): 591 volumes. Represents the typical size of existing volumetric datasets.
2. Base (4X-2X): 4,728 volumes. Each original volume is downsampled 4x and split into 8 sub-volumes.
3. Large (2X-4X): 37,824 volumes. Each original volume is downsampled 2x and split into 64 sub-volumes.

Classification Targets:
The dataset supports two hierarchical classification tasks:

• Coarse-grained: Classification of 25 cheese types (based on recipe parameters).
• Fine-grained: Classification of 149 individual samples (capturing subtle structural variations within the same cheese type).

3. Key Contributions

1. MozzaVID Dataset: The introduction of a large-scale, versatile volumetric dataset that bridges the gap between established 2D benchmarks and typical volumetric datasets. It offers three distinct scales to test model scalability.
2. Benchmarking Framework: A comprehensive evaluation of State-of-the-Art (SotA) architectures (ResNet, MobileNet, ConvNeXt, ViT, Swin) in both 2D (slice-based) and 3D configurations.
3. Structural Analysis via Deep Learning: Demonstration that deep learning classification can effectively quantify and map the variability of complex, disordered food microstructures, linking learned embeddings back to experimental design parameters.
4. Open Access: The dataset is publicly available, addressing the accessibility issues common in medical volumetric datasets.

4. Experimental Results

Model Performance:

• 3D vs. 2D: 3D models consistently outperformed their 2D counterparts (trained on single slices), highlighting the importance of volumetric representation. Notably, a smaller 3D dataset (Base) often outperformed a much larger 2D dataset (Large) for coarse classification.
• Architecture Comparison:
  • ResNet50 (3D) achieved the highest accuracy across most configurations.
  • MobileNetV2 showed stable performance, particularly in 2D.
  • Transformers (ViT, Swin) showed high variability; while Swin performed well on smaller datasets, ViT struggled with convergence on smaller volumetric data.
• Accuracy Scores:
  • Coarse-grained (25 classes): Reached 97.3% accuracy on the Large 3D instance.
  • Fine-grained (149 classes): Reached 93.5% accuracy on the Large 3D instance.
  • Performance dropped significantly on the "Small" and "Base" instances, confirming the difficulty of the task with limited data.

Representation Learning:

• UMAP Analysis: Visualizing the latent space of the best-performing model revealed distinct clustering of cheese types.
• Correlation with Metadata: The learned clusters aligned with the experimental design parameters (temperature, rotor speed, additives) visualized via PCA. This confirms that the models learned meaningful structural features rather than just memorizing noise.
• Anisotropy: The models successfully distinguished between anisotropic structures (fibrous) and isotropic structures (Cagliata cheese), validating the dataset's utility for structural analysis.

5. Significance and Impact

• Advancing Volumetric Deep Learning: MozzaVID provides the first large-scale benchmark to rigorously test if volumetric-specific architectures outperform adapted 2D models. The results suggest that while 3D data is superior, current SotA models (especially Transformers) may still be over-optimized for 2D, indicating a need for specialized volumetric research.
• Food Science Applications: The dataset enables the quantitative analysis of food microstructure. Understanding the relationship between processing parameters and 3D structure is crucial for engineering sustainable, high-quality plant-based food alternatives.
• Generalizability: Due to the disordered, organic nature of mozzarella, the dataset serves as a proxy for other complex biological and medical volumetric data, offering a new standard for evaluating structural analysis models.

In conclusion, MozzaVID addresses the critical lack of volumetric benchmarks, demonstrating that deep learning can effectively decode complex food microstructures while providing a rigorous testbed for the next generation of 3D computer vision algorithms.