Abstract
Convolutional Neural Networks (CNNs) are typically designed for translation invariance, assuming features can appear anywhere in an image. In neuroimaging, however, spatial normalization to standard templates like MNI space ensures that specific coordinates correspond to the same anatomical structures across subjects. We propose the Patch Individual Filter (PIF) layer, a novel architecture component that leverages this spatial consistency. By training filters locally rather than globally, PIF layers improve classification accuracy and training efficiency, especially in low-data regimes and tasks involving localized pathology.
Introduction
In computer vision, the principle of translation invariance allows CNNs to recognize objects regardless of their position by sharing weights across the entire image. While effective for natural images, this approach may be suboptimal for spatially normalized medical imaging. In neuroimaging, pre-processing steps align scans to a common template, making high-level features (e.g., hippocampal atrophy or specific white matter lesions) anatomically fixed. Standard global convolutions ignore this prior knowledge, potentially wasting model capacity and introducing noise. This is particularly critical given the high dimensionality of MRI data and the often limited sample sizes available in clinical research.
Methodology: The Patch Individual Filter (PIF) Layer
The PIF layer is designed to replace global convolutions at higher levels of the network where features become more anatomically specific. The layer operates in three distinct stages:
- Split: The input feature map is partitioned into a grid of patches.
- Process: Each patch is processed independently by its own set of local convolutions. Weights are shared within each patch but are unique to that specific spatial location. This allows the network to learn region-specific features.
- Reassemble: The processed patches are concatenated back into a single feature map for subsequent layers.
The authors recommend using standard global convolutions for low-level features (like edges) and transitioning to PIF layers for higher-level abstractions that correspond to specific brain regions.
Experimental Evaluation
The PIF architecture was evaluated across three neuroimaging tasks:
- Sex Classification: UK Biobank dataset (n=1,854).
- Alzheimer’s Disease (AD) Detection: ADNI dataset (n=969).
- Multiple Sclerosis (MS) Detection: Private clinical dataset (n=147).
To assess robustness in data-scarce environments, experiments were also conducted using only 20% of the training data for the UK Biobank and ADNI cohorts.
Results and Discussion
The PIF model demonstrated significant advantages in training speed and accuracy, particularly when data was limited.
Performance in Low-Data Regimes
On the UK Biobank sex classification task with reduced data, the PIF model achieved a balanced accuracy of 78.11%, compared to 64.47% for the baseline CNN. This suggests that the PIF architecture acts as an effective form of "informed" regularization by embedding anatomical priors into the network structure.
Training Efficiency
PIF layers consistently reached peak performance in fewer iterations. For the full UK Biobank dataset, PIF converged in 33.6 iterations, nearly twice as fast as the baseline's 59.2 iterations. Similar trends were observed in the MS detection task.
Task-Specific Efficacy
The PIF model excelled in MS detection (80.92% vs. 75.04% accuracy), likely because MS is characterized by localized lesions that benefit from region-specific filters. Conversely, on the ADNI dataset, the baseline slightly outperformed PIF. This may be due to the global nature of Alzheimer’s-related brain atrophy, which might be better captured by global filters.
Conclusion
The Patch Individual Filter (PIF) layer offers a promising approach to optimizing CNNs for spatially normalized medical data. By exploiting anatomical consistency, PIF layers reduce noise and improve learning efficiency. Future work will focus on determining the optimal transition point from global to local convolutions and exploring various patch configurations for different clinical applications.
Summary updated