Introduction
Deep Learning, particularly Convolutional Neural Networks (CNNs), has become the standard for medical imaging tasks. However, these models suffer from two significant drawbacks: the "black box" nature of their decision-making process and their high demand for labeled data and computational resources. To address these issues, we introduce Projectron, a shallow neural network architecture that utilizes Radon projections instead of raw pixels for image classification. By leveraging classical signal processing techniques, Projectron offers a transparent and efficient alternative to deep architectures.
Methodology: The Projectron Architecture
The Projectron framework moves away from pixel-based learning by transforming 2D images into 1D structural signals. The architecture is divided into three primary stages:
- Projection Stage: The input image undergoes a global Radon transform, generating 12 projections at equidistant angles (every 15° from 0° to 180°). This process captures the global shape and structural characteristics of the image.
- Encoding Stage: These 1D signals are processed through a single layer of neurons with ReLU activation, followed by a Radial Basis Function (RBF) kernel layer. This mapping allows the network to handle complex patterns within a shallow structure.
- Classification Stage: A two-layer Multi-Layer Perceptron (MLP) with a softmax function produces the final class probabilities.
Unlike traditional methods that use Radon transforms as static descriptors, Projectron learns the weights for these projections during training, effectively bridging the gap between signal processing and modern machine learning.
Experimental Evaluation
The performance of Projectron was evaluated across five diverse datasets, ranging from general benchmarks to specialized medical imaging:
- MNIST: Handwritten digits (70,000 images).
- CT Emphysema: Lung tissue patches (168 images).
- Invasive Ductal Carcinoma (IDC): Breast cancer histopathology (277,000 patches).
- IRMA: Radiograph images across 193 classes (14,410 images).
- Pneumonia: Chest X-rays (5,863 images).
Results and Discussion
Projectron demonstrated significant advantages in medical classification tasks compared to standard MLPs using raw pixel inputs. Notably, it achieved higher accuracy on the IDC (78% vs. 71%), IRMA (69% vs. 47%), and Pneumonia (70% vs. 37%) datasets. While performance on MNIST and the small Emphysema dataset was comparable to standard models, the results highlight Projectron's efficacy in handling complex, structured medical data.
Furthermore, Projectron exhibited superior training efficiency. On the MNIST dataset, it required approximately 2.5 minutes to train, whereas a standard MLP took over 4 minutes, despite the latter having a simpler layer structure.
Interpretability through Global Features
A core strength of Projectron is its inherent interpretability. By converting images into 1D projections, the network's focus can be visualized as graphs. This allows clinicians to correlate specific anatomical features—such as the density variations in a chest X-ray—with the model's output.
Conclusion and Future Directions
Projectron proves that shallow networks, when paired with intelligent input representations like the Radon transform, can rival or exceed the performance of deeper models in specific domains. While current comparisons focused on shallow MLPs, future work will involve benchmarking against state-of-the-art CNNs and exploring "learnable" projection angles to further optimize diagnostic accuracy. The integration of Radon projections into deeper architectures also remains a promising avenue for combining global structural awareness with deep feature extraction.
Summary updated