Abstract
Automating the classification of medical images in biomedical literature is a significant challenge due to high intra-class variation and substantial inter-class similarity. This paper introduces Synergic Deep Learning (SDL), a novel framework that utilizes a dual-network architecture linked by a "synergic signal." This signal enables the model to learn from pairs of images, effectively distinguishing between confusingly similar modalities. Evaluated on the ImageCLEF2016 benchmark, the SDL model achieved a state-of-the-art accuracy of 86.58%, outperforming deeper standalone models and setting a new record for the challenge.
Introduction
The classification of medical literature images is hindered by two primary visual hurdles. First, intra-class variation occurs when images in the same category (e.g., Computerized Tomography) look vastly different because they scan different anatomical regions. Second, inter-class similarity occurs when images from different categories (e.g., a brain CT vs. a brain MRI) share identical anatomical structures, making modality identification difficult for standard Deep Convolutional Neural Networks (DCNNs).
Synergic Deep Learning Framework
The core innovation of this work is the SDL model, which moves beyond standard ensemble methods by breaking the independence of networks during the training phase.
Dual-DCNN and Synergic Signal
The architecture consists of two identical DCNNs (e.g., ResNet-50). During training, the model is fed pairs of images. The "Synergic Signal System" acts as an information bridge that checks if the pair belongs to the same category.
Corrective Feedback
If the two networks produce a "synergic error"—such as disagreeing on a pair or failing to recognize a match—this error is backpropagated to both networks. This forces the models to mutually learn better representations, specifically targeting the "confusing" samples that typically lead to misclassification. During the inference phase, the synergic signal is discarded, and the final prediction is made by averaging the likelihood probabilities from both DCNN components.
Evaluation and Results
The authors evaluated the SDL model using the ImageCLEF2016 Subfigure Classification Challenge dataset, which contains 30 distinct classes including 12 medical diagnostic modalities and 18 types of illustrations.
- State-of-the-Art Performance: The SDL model achieved a classification accuracy of 86.58% on the test set, surpassing the previous first-place solution (85.38%) which utilized a deeper ResNet-152 model.
- Synergic Boost: Ablation studies showed that the synergic signal improved the accuracy of individual ResNet-50 components by more than 2% compared to standard standalone training.
- Robustness: The model demonstrated significant improvements in F-scores across almost all 30 categories, particularly those with high inter-class similarity.
Discussion and Future Work
SDL provides a scalable, backbone-agnostic framework to boost DCNN performance by focusing on sample relationships. While effective, the simultaneous training of two networks increases computational overhead. Future research could investigate "hard-negative mining" to select the most confusing image pairs and applying the synergic framework to more modern architectures like Vision Transformers.
Summary updated