Maximum Intensity Projection (MIP)

The core mechanism of MIP involves casting parallel rays through a volumetric slab of DBT data. For each ray, the algorithm traverses every voxel along its path and selects only the maximum intensity value to display on the final 2D pixel. This computationally efficient process discards low-intensity soft tissue, effectively isolating hyper-attenuating structures such as microcalcifications and biopsy clips. Unlike averaging techniques, MIP preserves the full contrast of the brightest object, ensuring subtle calcifications are not diluted by surrounding fibroglandular tissue.

The diagnostic utility of MIP is highly dependent on the slab thickness—the number of consecutive DBT slices included in the projection. Key considerations include:

• Thin slabs (5-10mm): Provide high depth resolution, allowing precise localization of a calcification's z-axis position but may fragment cluster morphology.
• Thick slabs (20-40mm): Visualize entire calcification clusters in a single view, revealing ductal distribution patterns, but obscure exact depth information.
• Sliding slabs: Radiologists often scroll through overlapping MIP reconstructions to mentally map the 3D volume.

MIP directly addresses the tissue overlap problem inherent in 2D mammography. In dense breast parenchyma, overlapping fibroglandular tissue can obscure or mimic calcifications. By projecting only the brightest voxels from a 3D volume, MIP suppresses the structured noise of overlapping tissue. This dramatically increases the conspicuity of microcalcifications, making subtle amorphous or pleomorphic clusters visible. This is the primary reason MIP slabs are the standard hanging protocol for DBT interpretation, serving as a rapid triage map before scrolling through individual 1mm slices.

In modern CADe systems, MIP images serve as a critical input channel for deep learning models. A 2D convolutional neural network can process an MIP slab to generate calcification candidate regions with high sensitivity. This approach offers computational advantages:

• Reduced dimensionality: A single MIP replaces a stack of 50-80 slices, lowering model inference time.
• Feature aggregation: The projection naturally consolidates 3D cluster morphology into a 2D representation compatible with established 2D detection architectures.
• Hybrid strategies: Advanced systems often run a 2D detector on MIP slabs for initial candidate generation, then refine false positives using a lightweight 3D model on the source slices.

While powerful, MIP introduces specific artifacts that AI engineers must address:

• High-density summation: Overlapping benign calcifications or a calcified artery crossing a rib can project as a pseudo-lesion mimicking a malignant cluster.
• Depth ambiguity: A MIP image contains no explicit depth information; a calcification at the top of the slab is indistinguishable from one at the bottom. This necessitates correlation with the source DBT slices.
• Motion blur propagation: Patient motion during DBT acquisition creates high-intensity edge artifacts that are faithfully projected into the MIP, potentially generating false positive CADe marks. Preprocessing with artifact suppression algorithms is essential.

MIP slabs are the default initial view in DBT interpretation, forming the foundation of the synthetic 2D mammogram (C-View or V-Preview). The standard clinical workflow proceeds as follows:

• Step 1: Radiologist reviews the MIP/synthetic 2D image for architectural distortion and calcification clusters.
• Step 2: Suspicious regions are immediately correlated with the 1mm native DBT slices by scrolling to the corresponding depth.
• Step 3: CADe marks are typically overlaid on the MIP slab, providing a single summary view of all algorithmic detections. This integration makes the MIP the central interface for AI-assisted reading.