Robust Multi-Source Covid-19 Detection in CT Images
Authors: Asmita Yuki Pritha, Jason Xu, Daniel Ding, Justin Li, Aryana Hou, Xin Wang, Shu Hu
Organizations: Capstone School Dhaka, Bangladesh · Carmel High School, Carmel, Indiana, USA · Clarkstown High School South, West Nyack, New York, USA · University at Albany, State University of New York, Albany, New York, USA · Purdue University, West Lafayette, Indiana, USA
Deep learning models for COVID-19 detection from chest CT scans generally perform well when the training and test data originate from the same institution, but they often struggle when scans are drawn from multiple centres with differing scanners, imaging protocols, and patient populations. One key reason is that existing methods treat COVID-19 classification as the sole training objective, without accounting for the data source of each scan. As a result, the learned representations tend to be biased toward centres that contribute more training data. To address this, we propose a multi-task learning approach in which the model is trained to predict both the COVID-19 diagnosis and the originating data centre. The two tasks share an EfficientNet-B7 backbone, which encourages the feature extractor to learn representations that hold across all four participating centres. Since the training data is not evenly distributed across sources, we apply a logit-adjusted cross-entropy loss [1] to the source classification head to prevent underrepresented centres from being overlooked. Our pre-processing follows the SSFL framework with KDS [2], selecting eight representative slices per scan. Our method achieves an F1 score of 0.9098 and an AUC-ROC of 0.9647 on a validation set of 308 scans. The code is publicly available at https://github.com/Purdue-M2/-multisource-covid-ct.
Figures & tables
Figure 1: Representative chest CT slices from two different scans across the four contributing medical centres, illustrating differences in field of view, contrast, tissue brightness, and windowing range that constitute the domain shift a multi-source model must overcome.
Figure 2: Overview of the proposed pipeline. Raw CT scans from COVID and non-COVID cases undergo lung extraction and KDS sampling to yield eight slices per scan. These pass through a shared EfficientNet-B7 backbone and are aggregated via element-wise mean pooling into a single 2560-dimensional scan-level representation. Two task-specific heads operate on this vector. The COVID-19 detection head is trained with binary cross-entropy ( ℓce ) and the auxiliary source identification head with logit-adjusted cross-entropy ( ℓLA ). The combined objective ℓ=ℓce+γ⋅ℓLA governs training, while only the detection head is used at inference.
Figure 3: Data preprocessing pipeline. Raw CT scans for both COVID-19 and Non-COVID cases undergo lung extraction and KDS sampling to select 8 highly representative slices per scan.
Training
Validation
Source
COVID
Non-COVID
Total
COVID
Non-COVID
Total
0
165
163
328
45
45
90
1
165
165
330
45
45
90
2
165
165
330
38
45
83
3
69
165
234
0
45
45
Total
564
658
1,222
128
180
308
Table 1: Per-source distribution of scans in the training and validation sets. Source 3 contains no COVID-19 cases in the validation partition.
Configuration
γ
Accuracy
F1 ↑
AUC ↑
Sensitivity
Specificity
Final Score ↑
Baseline (BCE)
–
0.9091
0.8915
0.9627
0.8984
0.9167
0.8008
Multi-task (CE)
0.1
0.9058
0.8889
0.9683
0.9062
0.9056
0.7996
Multi-task (CE)
0.2
0.9026
0.8846
0.9486
0.8984
0.9056
0.7923
Multi-task (CE)
0.5
0.8896
0.8692
0.9523
0.8828
0.8944
0.7904
Multi-task (CE)
1.0
0.9058
0.8930
0.9715
0.9453
0.8778
0.7942
Multi-task + LA
0.1
0.9123
0.8924
0.9675
0.8750
0.9389
0.7986
Table 2: Validation results across all configurations. Baseline = BCE only, no source head. Bold = best per column. ↑ = higher is better.
Method
Src 0
Src 1
Src 2
Src 3
Final ↑
Baseline (BCE)
0.9221
0.8776
0.9269
0.4767
0.8008
MT CE ( γ =0.1)
0.8888
0.9000
0.9389
0.4706
0.7996
MT + LA ( γ =0.5)
0.9555
0.8888
0.9756
0.4578
0.8194
Table 3: Per-source average F1 at each method’s best γ . Bold = best per source.
Figure 4: Per-source average F1 at each method’s best γ . The LA variant yields the largest gains on Source 0 (+3.3 percentage point) and Source 2 (+4.9 percentage point). Source 3 scores remain low across all methods due to the absence of COVID samples in the validation split.
Figure 5: Sensitivity to the source-loss weight γ . Final score of Multi-task + LA as γ varies in {0.1,0.2,0.5,1.0} . The dashed line marks the BCE-only baseline ( γ=0 ). Performance peaks at γ=0.5 (score =0.8194 ), where the source head and logit adjustment reinforce each other; both lower and higher weights degrade the score.
College of Computer Science and Artificial Intelligence, Shanghai Key Laboratory of Intelligent Information Processing, Fudan University · University of Oxford · The Hong Kong University of Science and Technology
VNUHCM University of Science, Vietnam National University, Vietnam · Ho Chi Minh University of Technology, Vietnam National University, Vietnam · The University of Alabama at Birmingham, United States