An Efficient Deep Neural Network to Classify Large 3D Images With Small Objects.

Area of Science:

• Medical Imaging AI
• Computational Pathology
• Machine Learning in Radiology

Background:

• 3D imaging offers superior anatomical detail for medical diagnosis but presents significant computational challenges for AI training due to large data volumes.
• Conventional AI models often downsample or project 3D images to 2D, potentially losing critical spatial information.
• Training AI on high-resolution 3D medical data requires substantial GPU memory and computational power, limiting accessibility and scalability.

Purpose of the Study:

• To introduce an efficient neural network, 3D Globally-Aware Multiple Instance Classifier (3D-GMIC), for classifying full-resolution 3D medical images.
• To demonstrate that 3D-GMIC can process large 3D datasets with significantly reduced computational resources compared to standard convolutional neural networks.
• To validate the model's ability to provide interpretable, pixel-level explanations for its predictions using saliency maps.

Main Methods:

• Developed a novel neural network architecture, 3D Globally-Aware Multiple Instance Classifier (3D-GMIC), designed for efficient processing of high-resolution 3D medical images.
• Trained and evaluated 3D-GMIC on a large dataset of 3D mammography, comparing its performance against 2D mammography (FFDM and synthetic) using image-level labels.
• Assessed computational efficiency by measuring GPU memory usage and computation time, and validated generalization on an external dataset (BCS-DBT).

Main Results:

• 3D-GMIC achieved an AUC of 0.831 for classifying malignant findings in 3D mammography, comparable to 2D methods (AUC 0.816-0.826).
• The model demonstrated significant reductions in GPU memory (77.98%-90.05%) and computation (91.23%-96.02%) compared to off-the-shelf convolutional neural networks.
• 3D-GMIC successfully identified small regions of interest in massive 3D images and generalized well to an external dataset, achieving an AUC of 0.848 on BCS-DBT.

Conclusions:

• 3D-GMIC offers an effective and computationally efficient solution for AI-driven classification of full-resolution 3D medical images.
• The model's ability to handle large 3D datasets without compromising accuracy or requiring extensive computational resources makes it a valuable tool for medical diagnosis.
• 3D-GMIC's interpretability through saliency maps and strong performance on external data highlight its potential for clinical application in breast cancer detection.