PHAROS-AIF-MIH

Abstract

We propose a deep learning framework for COVID-19 detection and disease classification from chest CT scans that integrates both 2.5D and 3D representations to capture complementary slice-level and volumetric information. The 2.5D branch processes multi-view CT slices (axial, coronal, sagittal) using a DINOv3 vision transformer, while the 3D branch employs a ResNet-18 architecture pretrained with Variance Risk Extrapolation (VREx) and further refined with supervised contrastive learning to improve cross-source robustness. Predictions from both branches are combined via logit-level ensembling.

Experiments on the PHAROS-AIF-MIH benchmark demonstrate the effectiveness of the proposed approach. On the test set, our method achieved runner-up in the Multi-Source COVID-19 Detection Challenge, with the best ensemble reaching a Macro F1-score of 0.751. For the Fair Disease Diagnosis Challenge, our approach ranked third place, achieving a best Macro F1-score of 0.633 with improved performance balance across genders. These results highlight the benefits of combining pretrained slice-based representations with volumetric modeling, as well as the importance of ensemble strategies for improving robustness and fairness in multi-source medical imaging tasks.

Repository: https://github.com/HySonLab/PHAROS-AIF-MIH

Dataset Structure

After downloading the competition datasets, organize them as follows:

data/
    task_1/
        1st_challenge_test_set/test
    covid1/
    covid2/
    non-covid1/
    non-covid2/
    non-covid3/
    Validation/val/
        covid/
        non-covid/
    train_covid.csv
    train_non_covid.csv
    validation_non_covid.csv

    task_2/
        2nd_challenge_test_set/test_for_participants
        train1/
            A/
            G/
        train2/
            covid/
            normal/
        Validation/val/
            A/
            G/
            covid/
            normal/

Preprocessing

task1_preprocess.py and task2_preprocess.py reconstruct 3D CT volumes from axial slices.

Processing steps:

• remove duplicate slices
• stack slices into a 3D volume
• resize to 128 × 128 × 128
• apply 3D Gaussian denoising and mask sharpening
• normalize to grayscale [0, 255]

The reconstructed volume enables extraction of coronal and sagittal views.

make_task1_csv.py and make_task2_csv.py generate CSV files indexing volumes and labels.

3D ResNet-18 Approach

Training Pipeline

Stage	Objective	Loss	Details
VREx Pretraining	Domain generalization	Cross-entropy + variance of domain losses	Domains: data source (Task 1) or gender (Task 2); 5 epochs; LR = 1e-4
Task Fine-tuning	Task-specific learning	Cross-entropy + supervised contrastive loss	MixUp (α = 0.4); 20 epochs; LR = 1e-5

Results

Task 1 (COVID detection)

Metric	Value
Accuracy	87.01%
Macro F1	0.7648

Per-source F1

Source	F1
Source 0	0.8630
Source 1	0.8408
Source 2	0.4828
Source 3	0.8725

Task 2 (A, G, COVID, Normal)

Metric	Value
Accuracy	76.77%
Macro F1	0.6677

2.5D DINOv3 Approach

Training Strategy

Stage	Description
Stage 1	Train classification head with frozen backbone
Stage 2	Partially unfreeze transformer layers
Stage 3	End-to-end fine-tuning

Results

Task 1

Metric	Value
Accuracy	93.51%
Macro F1	0.8221

Per-source F1

Source	F1
Source 0	0.9430
Source 1	0.9431
Source 2	0.4828
Source 3	0.9194

Task 2

Metric	Value
Accuracy	76.77%
Macro F1	0.7229

Ensemble

Predictions from the 2.5D and 3D models are combined using logit-level averaging.

Task	Accuracy	Macro F1
Task 1 (COVID detection)	95.13%	0.8321

Test set results

Multi-Source COVID-19 Detection (Task 1)

Method	Macro F1	H1	H2	H3	H4
2.5D	0.741	0.910	0.691	0.493	0.869
3D	0.699	0.825	0.608	0.495	0.868
Ensemble (0.3)	0.739	0.895	0.673	0.496	0.892
Ensemble (0.5)	0.751	0.914	0.712	0.495	0.883
Ensemble (0.7)	0.744	0.912	0.696	0.495	0.873

Fair Disease Diagnosis (Task 2)

Method	Macro F1	Female	Male
2.5D	0.555	0.709	0.402
3D	0.572	0.756	0.388
Ensemble (0.3)	0.568	0.732	0.403
Ensemble (0.5)	0.561	0.718	0.403
Ensemble (0.7)	0.633	0.709	0.557

If our work is useful, please cite us!

@misc{yang2026halfway3densembling25d,
      title={Halfway to 3D: Ensembling 2.5D and 3D Models for Robust COVID-19 CT Diagnosis}, 
      author={Tuan-Anh Yang and Bao V. Q. Bui and Chanh-Quang Vo-Van and Truong-Son Hy},
      year={2026},
      eprint={2603.14832},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.14832}, 
}