Robust Multi-Source Covid-19 Detection in CT Images

1. Problem Statement

The paper addresses the critical challenge of domain shift in deep learning models for COVID-19 detection using chest CT scans.

• The Core Issue: While models perform well when training and testing data come from the same institution, they often fail when deployed across multiple centers. This is due to variations in scanner hardware, imaging protocols, slice thickness, and patient demographics.
• Data Imbalance: In multi-source datasets, data contribution is rarely uniform. Some centers provide significantly more training data than others. Standard training methods cause the model to overfit to the features of the data-rich centers (majority bias), leading to poor generalization on underrepresented centers.
• Current Limitations: Existing methods typically treat COVID-19 classification as a single objective. They do not explicitly account for the data source, allowing the model to learn "acquisition artifacts" (e.g., specific scanner brightness or reconstruction kernels) rather than true pathological features, resulting in biased decision boundaries.

2. Methodology

The authors propose a Multi-Task Learning (MTL) framework designed to learn source-invariant representations. The pipeline consists of three main components:

A. Preprocessing (SSFL + KDS)

To handle heterogeneous inputs from four different hospitals, the authors adopt the SSFL (Slice Sampling for Fast Learning) framework with Kernel-Density-based Slice Sampling (KDS):

• Lung Extraction: A 3×3 minimum filter removes noise, followed by binarization and morphological hole filling to isolate the lung region. Images are resized to 256×256.
• Slice Selection: Instead of uniform subsampling (which might miss diagnostic regions in short scans or waste computation on redundant slices in long scans), KDS fits a Gaussian kernel density estimate to the lung area distribution. It selects 8 representative slices per scan based on percentile intervals, ensuring anatomical coverage regardless of the original scan length.

B. Multi-Task Architecture

The model utilizes a shared EfficientNet-B7 backbone (pretrained on ImageNet) to extract features.

• Input: The 8 selected slices pass independently through the backbone, producing 8 feature vectors (2560 dimensions each).
• Aggregation: These vectors are aggregated via element-wise mean pooling into a single scan-level representation.
• Dual Heads:
  1. Primary Task (Diagnosis): A binary classification head (Sigmoid activation) predicts COVID-19 status.
  2. Auxiliary Task (Source Identification): A 4-class classification head (Softmax activation) predicts the originating medical center.
• Rationale: By forcing the shared backbone to predict the source, the model is compelled to disentangle "disease features" from "acquisition artifacts." The goal is for the backbone to learn features that are useful for diagnosis regardless of the source, effectively suppressing institutional confounders.

C. Logit-Adjusted Loss Function

To address the imbalance in data contribution (three centers contribute ~330 scans, one contributes ~234), the authors apply Logit-Adjusted Cross-Entropy (LA) to the source classification head.

• Mechanism: A correction term based on the log-frequency of each source ($\log(p(d))$) is added to the logits before the softmax operation.
• Effect: This offsets the bias toward majority classes. It ensures that the model cannot simply "coast" by predicting the most common center; it must learn to distinguish underrepresented sources, thereby forcing the shared encoder to develop a truly source-invariant feature space.
• Total Objective: The total loss is a weighted sum: $\mathcal{L} = \mathcal{L}_{CE} + \gamma \cdot \mathcal{L}_{LA}$, where $\gamma$ controls the influence of the source task.

3. Key Contributions

1. Multi-Task Formulation: The paper reframes multi-source detection as a multi-task problem, pairing diagnosis with source identification to encourage the learning of transferable features across institutional boundaries.
2. Logit-Adjusted Supervision: The authors demonstrate that standard multi-task learning fails if the auxiliary task is not corrected for class imbalance. They introduce logit-adjusted cross-entropy for the source head, which prevents the model from being dominated by data-rich centers.
3. Empirical Validation: The study provides a rigorous ablation study showing that logit adjustment improves the final score by 1.86 percentage points over a single-task baseline and 1.98 percentage points over a standard multi-task variant.

4. Experimental Results

• Dataset: 3,020 CT scans from 4 distinct medical centers (1,222 for training, 308 for validation).
• Metrics: F1 Score, AUC-ROC, and a specific Final Score (mean of per-source F1 averages) designed to penalize models that perform poorly on specific centers.
• Performance:
  • Baseline (Single-task): Final Score = 0.8008.
  • Multi-Task (Standard CE): Performed worse than the baseline (max 0.7996), confirming that uncorrected source supervision introduces bias.
  • Multi-Task + Logit Adjustment (Optimal $\gamma=0.5$): Achieved a Final Score of 0.8194, an F1 of 0.9098, and an AUC-ROC of 0.9647.
• Per-Source Analysis: The proposed method significantly improved performance on underrepresented or difficult sources (Source 0 and Source 2 saw gains of +3.34% and +4.87% respectively), proving the model's ability to generalize across domains.
• Hyperparameter Sensitivity: The weighting parameter $\gamma$ was critical. $\gamma=0.5$ provided the best balance; $\gamma=1.0$ caused the model to overfit to source identity, degrading disease detection performance.

5. Significance and Conclusion

This paper highlights that multi-task learning is not a silver bullet; its success depends heavily on how the auxiliary task is formulated.

• Key Insight: Simply adding a source classification head is insufficient and can be harmful if the data is imbalanced. The auxiliary loss must be explicitly designed to counteract the data distribution (via logit adjustment) to force the model to learn domain-invariant features.
• Clinical Impact: The proposed framework offers a robust solution for deploying AI in real-world, multi-center hospital settings where scanner protocols and patient demographics vary, ensuring that diagnostic accuracy does not degrade for patients from underrepresented institutions.
• Future Directions: The authors suggest relaxing the dependency on explicit source labels (using unsupervised domain discovery) and incorporating patient-level fairness objectives to further address demographic biases.

The code for the framework is publicly available, promoting reproducibility in medical AI research.