[Paper Review] A Heterogeneous Ensemble for Multi-Center COVID-19 Classification from Chest CT Scans

Aadit Nilay, Bhavesh Thapar+2 more|arXiv (Cornell University)|Mar 15, 2026

COVID-19 diagnosis using AI0 citations

TL;DR

The paper introduces a heterogeneous ensemble of nine models spanning three inference paradigms to classify COVID-19 from chest CT scans across multiple centers, achieving a macro F1 of 0.9280 on four hospital centers and outperforming the best single model by 0.031.

ABSTRACT

The COVID-19 pandemic exposed critical limitations in diagnostic workflows: RT-PCR tests suffer from slow turnaround times and high false-negative rates, while CT-based screening offers faster complementary diagnosis but requires expert radiological interpretation. Deploying automated CT analysis across multiple hospital centres introduces further challenges, as differences in scanner hardware, acquisition protocols, and patient populations cause substantial domain shift that degrades single-model performance. To address these challenges, we present a heterogeneous ensemble of nine models spanning three inference paradigms: (1) a self-supervised DINOv2 Vision Transformer with slice-level sigmoid aggregation, (2) a RadImageNet-pretrained DenseNet-121 with slice-level sigmoid averaging, and (3) seven Gated Attention Multiple Instance Learning models using EfficientNet-B3, ConvNeXt-Tiny, and EfficientNetV2-S backbones with scan-level softmax classification. Ensemble diversity is further enhanced through random-seed variation and Stochastic Weight Averaging. We address severe overfitting, reducing the validation-to-training loss ratio from 35x to less than 3x, through a combination of Focal Loss, embedding-level Mixup, and domain-aware augmentation. Model outputs are fused via score-weighted probability averaging and calibrated with per-source threshold optimization. The final ensemble achieves an average macro F1 of 0.9280 across four hospital centres, outperforming the best single model (F1=0.8969) by +0.031, demonstrating that heterogeneous architectures combined with source-aware calibration are essential for robust multi-site medical image classification.

Motivation & Objective

• Address domain shift across multiple hospital centers caused by scanner, protocol, and patient population differences.
• Develop a robust heterogeneous ensemble to improve COVID-19 classification from chest CTs.
• Mitigate overfitting and improve generalization with advanced training and augmentation strategies.
• Calibrate model outputs using per-source thresholds to enhance multi-site performance.

Proposed method

• Use three inference paradigms: (1) self-supervised DINOv2 Vision Transformer with slice-level sigmoid aggregation; (2) RadImageNet-pretrained DenseNet-121 with slice-level sigmoid averaging; (3) seven Gated Attention MIL models with backbones EfficientNet-B3, ConvNeXt-Tiny, and EfficientNetV2-S with scan-level softmax classification.
• Enhance ensemble diversity via random-seed variation and Stochastic Weight Averaging.
• Address overfitting with Focal Loss, embedding-level Mixup, and domain-aware augmentation.
• Fuse model outputs with score-weighted probability averaging and calibrate using per-source threshold optimization.

Experimental results

Research questions

• RQ1 Can a heterogeneous ensemble improve multi-center COVID-19 CT classification under domain shifts?
• RQ2 Does source-aware calibration (per-center thresholds) outperform uniform calibration in multi-site evaluation?
• RQ3 How do diverse architectures and training strategies contribute to robustness against center-specific variations?

Key findings

• The ensemble achieves an average macro F1 of 0.9280 across four hospital centers.
• The ensemble outperforms the best single model (F1 = 0.8969) by 0.031.
• Overfitting is reduced, with the validation-to-training loss ratio dropping from 35x to less than 3x.
• Diversity from multiple architectures and training paradigms, plus calibration, is key for robust multi-site performance.