Abstract
Feature extraction is a very crucial task in image and pixel (voxel) classification and regression in biomedical image modeling. In this work we present a machine learning based feature extraction scheme based on inception models for pixel classification tasks. We extract features under multi-scale and multi-layer schemes through convolutional operators. Layers of Fully Convolutional Network are later stacked on this feature extraction layers and trained end-to-end for the purpose of classification. We test our model on the DRIVE and STARE public data sets for the purpose of segmentation and centerline detection and it out performs most existing hand crafted or deterministic feature schemes found in literature. We achieve an average maximum Dice of 0.85 on the DRIVE data set which out performs the scores from the second human annotator of this data set. We also achieve an average maximum Dice of 0.85 and kappa of 0.84 on the STARE data set. Though these datasets are mainly 2-D we also propose ways of extending this feature extraction scheme to handle 3-D datasets.
Introduction
In biomedical imaging, identifying structures like blood vessels (segmentation) and their paths (centerline prediction) is vital for diagnosis. Traditionally, researchers relied on hand-crafted features—mathematical descriptors like the "Hessian" or "vesselness" filters—to highlight tubular structures. While effective, these are limited by the designer's intuition and often fail to capture the full complexity of biological variation.
Modern Deep Learning, specifically Convolutional Neural Networks (CNNs), has revolutionized this field. However, most standard architectures are designed for whole-image classification and utilize "pooling" layers that shrink image dimensions. This reduction in spatial resolution is detrimental to pixel-level tasks like vessel segmentation, where fine-grained detail of tiny capillaries must be preserved. Deep-FExt addresses this by building a deep, powerful feature extractor that remains "pixel-perfect" while capturing features at multiple scales simultaneously.
Methodology
The proposed Deep-FExt (Deep Feature Extraction) framework modifies the Inception architecture specifically for high-precision pixel-wise tasks.
Key Architectural Innovations:
- Pixel-wise Inception Layers: Unlike standard Inception modules, Deep-FExt removes all pooling operations within the feature extraction stages. This ensures every convolution maintains the original input resolution, allowing for precise localization.
- Factorized Convolutions: To maintain computational efficiency while increasing model depth, the authors utilize "mini-networks." For instance, a single 5x5 convolution is replaced by two stacked 3x3 layers. This reduction in parameters allows for more non-linearities and a deeper hierarchy of features.
- Multi-Scale Integration: Each layer extracts features at various scales using different kernel sizes (e.g., 3, 5, 7, and 9). This allows the network to recognize vessels of varying diameters.
- The Feature Mesh: After five extraction layers, each pixel is represented by 100 distinct features. These are organized into a 10x10 "feature mesh," which is then processed by a final 3-layer CNN for classification.
Experimental Setup
The model was evaluated on two gold-standard public datasets for retinal imaging:
- DRIVE: 40 high-resolution images of the human retina with manual annotations.
- STARE: 20 images used to test the robustness of vessel segmentation algorithms.
To test the model's ability to predict the vessel's core path, the authors generated centerline annotations by "skeletonizing" the existing vessel masks from these datasets.
Key Insights and Ideas
- Dimensionality Extension: While most AI architectures focus on data compression, pixel-wise segmentation benefits from extension. By transforming one RGB pixel into 100 deep features, the model provides the classifier with a rich "contextual biography" of that pixel's neighborhood.
- Depth over Width: By factorizing large kernels into smaller ones, the network can learn hierarchical features—starting from simple edges and moving up to complex curvatures and vessel junctions—without an exponential increase in parameters.
- Surpassing the Human Baseline: A significant finding was that the model achieved a higher Dice score than the second human annotator on the DRIVE dataset, suggesting that the AI is more consistent than human experts in tracing fine, low-contrast vessels.
Key Results
- Segmentation Accuracy: On the DRIVE dataset, Deep-FExt achieved an average maximum Dice score of 0.85 and a Kappa of 0.84.
- Comparative Performance: The "Fusion" version of the model outperformed prominent architectures like DRIU and HED across Precision, Recall, and F1 metrics.
- Centerline Prediction: The model proved highly effective at finding the vessel core, achieving a Dice score of 0.77 on DRIVE, significantly higher than the human annotator's 0.69.
Conclusion and Future Work
Deep-FExt demonstrates that learned features can significantly outperform deterministic, hand-crafted filters in biomedical tasks. While currently focused on 2D retinal images, the architecture is designed to be extensible to 3D datasets by utilizing 3D convolutional operators. Future work may explore unsupervised feature extraction via autoencoders to mitigate the need for extensive manual labeling, which remains a bottleneck in medical AI development.
Summary updated