Abstract
This paper proposes a novel image feature set based on a principled information theoretic analysis of the convolutional neural network (CNN). The output of convolutional filters is modeled as a random variable conditioned on the object class and network filter bank. The conditional entropy (CENT) of filter outputs is shown in theory and experiments to be a highly compact and class-informative feature that can be computed from the CNN feature maps and used to obtain higher classification accuracy than the original CNN itself. Experiments involve three binary classification tasks using the 3D brain MRI data: Alzheimer's disease (AD) versus healthy controls (HC), young versus old age, and male versus female, where the area under the curve (AUC) values for the CENT feature classification (93.9%, 96.7%, and 71.9%) are significantly higher than the softmax output of the original CNN classifier trained for the task (81.6%, 79.4%, and 63.1%). A statistical analysis based on the Wilcoxon test identifies CENT features with significant links to brain labels, which could potentially serve as diagnostic biomarkers.
Introduction
While Convolutional Neural Networks (CNNs) have revolutionized image classification, they are frequently criticized as "black boxes," particularly in medical imaging where interpretability is paramount. Medical datasets, such as 3D brain MRIs, often suffer from data scarcity, making it difficult for deep learning models to generalize effectively. Traditional radiomics offers interpretability through hand-crafted features (e.g., texture or shape) but often lacks the predictive power of deep learning.
This research introduces Conditional Entropy (CENT) features to bridge this gap. By treating a trained CNN as an information-theoretic feature extractor, the authors capture highly compact and discriminative information from intermediate layers. This approach suggests that the intermediate feature maps contain a wealth of class-specific information that is often discarded or underutilized by the final softmax layer of a standard network.
Methods
The proposed approach transitions from end-to-end CNN classification to a hybrid "deep radiomics" framework. The process involves four primary stages:
- CNN Training: A relatively simple 3D CNN, consisting of two convolutional layers and one fully connected layer, is trained on a subset of brain MRI data to establish a baseline feature extractor.
- Information Flow Analysis: The output of convolutional filters is modeled as a random variable. The method analyzes the "information flow" through the network by calculating the entropy of these filter outputs conditioned on the object class and the specific filter bank.
- CENT Feature Extraction: This process generates a "compact code" consisting of only 21 features. These features represent the uncertainty reduction provided by specific filters regarding the class label (e.g., Alzheimer's vs. Healthy).
- Statistical Classification: These 21 CENT features are fed into a Random Forest (RF) classifier. This shifts the final decision-making from the CNN's softmax layer to a more robust statistical model capable of handling smaller feature sets.
Datasets and Preprocessing
The study utilized two major public neuroimaging repositories:
- OASIS-1 (Open Access Series of Imaging Studies): Used for Alzheimer’s Disease (AD) vs. Healthy Controls (198 subjects) and Young vs. Old (329 subjects) classification.
- ABIDE I (Autism Brain Imaging Data Exchange): Used for a gender classification task (Male vs. Female) involving 198 healthy subjects.
All images underwent standard preprocessing, including motion correction, normalization, and skull-stripping using the Freesurfer tool, ensuring consistency across the various classification tasks.
Results
The CENT-based approach consistently and significantly outperformed the standard CNN softmax output across all three tasks, demonstrating the power of extracting information-theoretic features from intermediate layers:
| Task | CENT + RF (AUC) | Standard CNN Softmax (AUC) | Improvement |
|---|---|---|---|
| AD vs. Healthy | 93.9% | 81.6% | +12.3% |
| Young vs. Old | 96.7% | 79.4% | +17.3% |
| Male vs. Female | 71.9% | 63.1% | +8.8% |
A statistical analysis using the Wilcoxon test confirmed that these CENT features have significant links to biological brain labels. This suggests that the features are not merely statistical artifacts but represent meaningful neuroanatomical biomarkers.
Discussion and Conclusion
The results support the Information Bottleneck theory, which posits that as data moves through a CNN, the network acts as a bottleneck, progressively reducing the entropy of the representation to focus only on class-relevant information. Interestingly, the authors found that features from the second convolutional layer were often more informative for Alzheimer's detection than the first layer, suggesting that mid-level abstractions are key for identifying complex neuroanatomical changes.
The CENT method successfully compresses the massive amount of data in a 3D MRI into just 21 highly potent numbers without losing predictive power. While the current study focused on MRI, future work is needed to test the generalizability of CENT features on other modalities like CT or PET scans. Furthermore, applying CENT in a spatially localized fashion could help clinicians identify the specific brain regions most affected by a disease, significantly increasing the clinical utility of the model.
Summary updated