TL;DR
This paper introduces a novel 2D convolutional neural network (CNN) approach for predicting fetal brain age using multiplanar MRI slices. By leveraging information from orthogonal planes and aggregating predictions using a mode-based strategy, the method achieves highly accurate brain age estimation with a mean absolute error of just 0.125 weeks, outperforming existing 2D multi-channel and 3D CNN approaches.
What Problem Does This Paper Address?
Accurately determining fetal brain age from MRI is crucial for identifying potential developmental abnormalities and predicting future outcomes. Traditional methods often rely on manual extraction of quantitative features (like cortical gyrification or sulcal depth), which is time-consuming and prone to error. While deep learning has shown promise, existing approaches using 3D CNNs or multi-channel 2D CNNs can be computationally intensive and sensitive to inter-slice motion artifacts inherent in fetal MRI. This study aims to bridge this gap by developing a more efficient and robust method that utilizes multiplanar 2D slices without complex preprocessing, thereby reducing dimensionality and improving accuracy.
Methods
The core of this work is a 2D single-channel CNN, based on the ResNet101V2 architecture. Key methodological innovations include:
- Multiplanar Slice Input: Instead of using a single slice or a full 3D volume, the CNN takes multiple slices from different orthogonal planes (sagittal, coronal, axial) as input. This allows the model to capture 3D spatial information without the computational burden of full 3D processing.
- Test-Time Augmentation (TTA): To enhance robustness and reduce prediction error, TTA is applied. This involves generating multiple augmented versions of each slice (e.g., flips, rotations, shifts) and averaging their predictions.
- Mode-Based Aggregation: For each fetus, multiple age predictions are generated from different slices. To obtain a single, reliable brain age estimate, the authors compare three central tendency measures: mean, median, and mode. They found that the mode (most frequent prediction) with a 0.2-week class interval yielded the lowest error.
- 2D U-Net for Brain Masking: An in-house 2D U-Net model was used for automatic brain segmentation, simplifying preprocessing.
Datasets
- Primary Dataset: 220 healthy fetuses from Boston Children's Hospital (BCH) with gestational ages ranging from 15.9 to 38.7 weeks.
- External Validation Dataset: 21 healthy fetuses from Tufts Medical Center (TMC), acquired on a different scanner (1.5T vs. 3T) and with slightly different acquisition parameters. This dataset was used to assess the generalizability of the proposed method.
Key Insights and Ideas
- Multiplanar Slices are Key: The study strongly suggests that utilizing multiple orthogonal 2D slices is a powerful strategy for capturing 3D brain information without the full complexity of 3D CNNs. This offers a sweet spot between single-slice limitations and 3D CNN demands.
- Mode Aggregation Outperforms Mean/Median: The choice of aggregation strategy for multiple slice predictions significantly impacts accuracy. The mode, by identifying the most frequent prediction, acts as a robust outlier suppressor, leading to better results than simple averaging.
- Ventricles and Cortex are Critical Predictors: Saliency maps reveal that the ventricles and cerebral cortex are the most important anatomical regions for predicting fetal brain age. This aligns with known developmental processes, lending biological plausibility to the model's findings.
- Brain Size Matters: The study confirms that overall brain size is a significant factor in age prediction, which is expected given the strong correlation between brain growth and gestational age.
Key Results
- High Accuracy on Primary Dataset: The proposed 2D single-channel CNN with multiplanar slices achieved a mean absolute error (MAE) of 0.125 weeks (0.875 days) on the internal dataset.
- Superior Performance: This method significantly outperformed:
- Single-slice and single-stack approaches (lower MAE and variance).
- 2D multi-channel CNN (stack-wise MAE of 0.304 weeks vs. 0.979 weeks).
- 3D CNN (stack-wise MAE of 0.304 weeks vs. 1.114 weeks).
- External Validation Success: On the external TMC dataset, the method achieved an MAE of 0.508 weeks, demonstrating good generalizability despite differences in scanner and acquisition parameters.
- Saliency Map Insights: The cortex and ventricles were identified as the primary contributors to age prediction, with consistent high saliency values.
Limitations and Future Work
- Overfitting Concerns: The authors acknowledge that the high performance on the internal dataset might be due to overfitting, as the model was trained on single-site data. While TTA and architectural choices (dropout, early stopping) were used to mitigate this, training on larger, multi-site datasets is recommended for improved generalizability.
- Batch Size Limitations for 3D CNN: The 3D CNN experiments were limited by GPU memory, resulting in a smaller batch size than the 2D CNN. This could have impacted its performance, and more hardware resources would be needed for a fair comparison.
- GA Estimation Sensitivity: The accuracy of the predicted brain age is inherently dependent on the accuracy of the ground truth gestational age (GA), which can have systematic errors.
- Clinical Relevance: The authors suggest applying the method to clinically abnormal fetuses to evaluate its clinical utility and investigating the relationship between predicted age differences (PADs) and neurodevelopmental outcomes.
- Demographic Factors: Future work could explore the influence of demographic factors (e.g., parental race/ethnicity) on age prediction.
Summary updated