Technical Summary: Taxlifier

Problem Statement

Accurate and efficient classification of thoracic diseases in chest X-ray (CXR) images is critical for timely diagnosis. However, automated systems face significant challenges due to the presence of multiple pathologies with overlapping visual characteristics and the complex, hierarchical nature of medical taxonomies. Conventional multi-label classification methods, such as One-vs-All (OVA), often treat labels independently, ignoring dependencies between classes. This "flat classification" approach fails to capture the conditional relationships inherent in medical diagnoses (e.g., a child pathology like consolidation implies the presence of a parent pathology like lung opacity). Furthermore, leaf nodes in disease hierarchies often suffer from a scarcity of positive examples, leading to negative class bias in standard models. While hierarchical classification has been explored in text processing and genomics, there is a lack of research applying taxonomical information to multi-class problems in medical imaging.

Methodology

The paper proposes Taxlifier, a framework introducing two novel hierarchical multi-label classification techniques designed to leverage disease taxonomy relationships. Both methods utilize a pre-defined graph structure $G = \{C, E\}$ representing parent-child relationships among thoracic pathologies.

1. Logit-Based Approach (Post-Processing)

This method is designed for scenarios where ground truth is unavailable or retraining is computationally prohibitive. It adjusts the predicted probabilities of a pre-trained neural network (e.g., DenseNet121) during inference without modifying the model architecture.

• Mechanism: The approach calculates a conditional predicted probability for a child class $c_k$ based on the logit output of its parent class $c_j$.
• Formula: The updated probability $\hat{p}^{(i)}_k$ is derived by modifying the original logit $q^{(i)}_k$ with a weighted contribution from the parent's logit $q^{(i)}_j$:
  $$\hat{p}^{(i)}_k = \frac{1}{1 + \exp(-(q^{(i)}_k + \alpha_k q^{(i)}_j))}$$
  where $\alpha_k$ is a hyperparameter controlling the influence of the parent class.
• Optimization: Hyperparameters $\{\alpha_k\}$ are tuned using a Tree-structured Parzen Estimator (TPE) to minimize the objective function $1 - \text{AUC}$ on a validation set.

2. Loss-Based Approach (Training-Time)

This method integrates hierarchical information directly into the model's optimization process via a novel regularization term.

• Mechanism: Instead of adjusting outputs post-hoc, this approach penalizes misalignment between predictions and the label taxonomy during training. It updates the loss value of a child class based on the loss values of its parent classes.
• Formula: The updated loss $\hat{l}^{(i)}_k$ is calculated as:
  $$\hat{l}^{(i)}_k = l^{(i)}_k \cdot H(k|j)$$
  where $H(k|j) = \alpha_k l^{(i)}_j + \beta_k$. Here, $l^{(i)}_j$ is the loss of the parent class, and $\alpha_k, \beta_k$ are hyperparameters balancing the influence of the parent's loss.
• Probability Update: To avoid re-running the full optimization, the paper derives a method to update predicted probabilities directly from the modified loss values using the gradient of the binary cross-entropy and sigmoid functions:
  $$\hat{p}^{(i)}_k = \begin{cases} \exp(-\hat{l}^{(i)}_k) & \text{if } p^{(i)}_k > \theta_k \text{ and } p^{(i)}_j > \theta_j \\ 1 - \exp(-\hat{l}^{(i)}_k) & \text{otherwise} \end{cases}$$
• Optimization: Hyperparameters $\{\alpha_k, \beta_k\}$ are optimized using TPE to maximize AUC.

Experimental Setup

The techniques were evaluated on three large-scale public datasets: CheXpert (224,316 images), PADCHEST (160,000 images), and NIH (112,120 images). A unified taxonomy was constructed by merging existing taxonomies from these datasets, focusing on 18 common pathologies. The baseline model was a DenseNet121 fine-tuned on these datasets. Performance was measured using Accuracy, AUC, F1-score, and comprehensive statistical metrics (Cohen's Kappa, t-statistic, Bayes Factor).

Key Contributions

1. Loss-Based Technique: Introduces a novel hierarchical regularization term that directly incorporates label hierarchy into the training loss function, penalizing misalignment with the taxonomy.
2. Logit-Based Technique: Proposes a computationally efficient post-processing method to adjust neural network logits based on class relationships, enabling the use of taxonomical knowledge without extensive retraining.
3. Comprehensive Evaluation: Demonstrates consistent performance improvements across three diverse, large-scale CXR datasets.
4. Interpretability: The hierarchical structure provides a more interpretable output, allowing clinicians to understand predictions at varying levels of granularity (parent vs. child classes).

Results

The experimental results indicate significant improvements over the baseline method across various pathologies, particularly for child nodes in the hierarchy:

• Accuracy: The logit-based and loss-based methods improved accuracy by 12% and 11%, respectively, compared to the baseline.
• AUC: Improvements of 13% (logit-based) and 10% (loss-based) were observed.
• F1-Score: The methods yielded substantial gains, with improvements of 24% (logit-based) and 12% (loss-based).
• Statistical Significance: Statistical analysis (t-tests, Bayes Factors) confirmed that the improvements, particularly for the logit-based method, were statistically significant across most pathologies (e.g., Atelectasis, Edema, Cardiomegaly). The loss-based method showed higher variance but achieved superior performance in specific cases (e.g., Lung Lesion F1 score increased from 0.094 to 0.982).
• Parent Classes: As expected, performance for parent classes (e.g., Lung Opacity, Enlarged Cardiomediastinum) remained consistent with the baseline, as the techniques specifically target the refinement of child class predictions based on parent relationships.

Significance and Claims

The paper claims that integrating domain-specific hierarchical knowledge significantly enhances both the accuracy and interpretability of computer-aided diagnosis (CAD) systems for chest radiography. By moving away from "flat" classification, the proposed methods address the negative class bias common in hierarchical medical data and provide a more robust framework for detecting complex, multi-label thoracic diseases.

The authors emphasize that the logit-based approach offers a practical advantage for real-world deployment due to its computational efficiency and lack of need for model retraining, making it adaptable to scenarios with limited resources. The loss-based approach offers a deeper integration of hierarchy during the learning phase.

The study concludes that these techniques represent a valuable contribution to multi-label classification in medical imaging, potentially reducing diagnostic errors and streamlining clinical workflows. However, the authors modestly acknowledge limitations, noting that the effectiveness relies on the quality of the taxonomy and dataset labeling, and that applying these methods to new domains requires the development of specific taxonomical structures by domain experts.