[Paper Review] Cov3d: Detection of the presence and severity of COVID-19 from CT scans using 3D ResNets

Robert Turnbull|arXiv (Cornell University)|Jul 5, 2022

COVID-19 diagnosis using AI5 citations

TL;DR

Cov3d proposes a 3D ResNet-based deep learning model to detect the presence and severity of COVID-19 from chest CT scans using the COV19-CT-DB dataset. It achieves a macro F1 score of 0.9476 for presence detection and 0.7552 for severity classification, outperforming the MIA-COV19D 2022 competition baseline.

ABSTRACT

Deep learning has been used to assist in the analysis of medical imaging. One such use is the classification of Computed Tomography (CT) scans when detecting for COVID-19 in subjects. This paper presents Cov3d, a three dimensional convolutional neural network for detecting the presence and severity of COVID19 from chest CT scans. Trained on the COV19-CT-DB dataset with human expert annotations, it achieves a macro f1 score of 0.9476 on the validation set for the task of detecting the presence of COVID19. For the task of classifying the severity of COVID19, it achieves a macro f1 score of 0.7552. Both results improve on the baseline results of the `AI-enabled Medical Image Analysis Workshop and Covid-19 Diagnosis Competition' (MIA-COV19D) in 2022.

Motivation & Objective

• To develop a deep learning model capable of detecting both the presence and severity of COVID-19 in chest CT scans.
• To improve upon the baseline results of the MIA-COV19D 2022 competition, which focused on dual-task classification from CT imaging.
• To leverage 3D convolutional neural networks to capture spatial context across axial, sagittal, and coronal planes in volumetric CT data.
• To address the challenge of variable CT scan depth and resolution through standardized preprocessing and data augmentation.
• To optimize model performance using a weighted multi-task loss function that prioritizes detection accuracy while maintaining severity classification performance.

Proposed method

• Cov3d employs a 3D ResNet-18 architecture adapted from video-pretrained models to process volumetric CT data.
• Input CT scans are preprocessed by resizing slices to 128², 256², or 320² pixels via bicubic interpolation and then resampling depth using 1D cubic interpolation to match the width/height.
• The model is trained using a custom loss function combining cross-entropy for both detection and severity classification, with a hyperparameter λ to balance the two tasks.
• Data augmentation includes random sagittal flipping during training, with test-time inference averaging predictions from original and flipped inputs.
• Weight decay of 10⁻⁵ and Adam optimization with cosine annealing are used to regularize training and improve generalization.
• Model weights are saved based on the highest macro F1 score achieved on the validation set for each task.

Experimental results

Research questions

• RQ1 Can a 3D convolutional neural network effectively detect the presence of COVID-19 in chest CT scans with high accuracy?
• RQ2 To what extent can a single deep learning model simultaneously predict both the presence and severity of COVID-19 from CT imaging?
• RQ3 How does preprocessing CT scans with variable depth and resolution impact model performance and training stability?
• RQ4 What hyperparameter settings, including loss weighting and data augmentation, yield optimal performance for dual-task classification in medical imaging?
• RQ5 Does the use of 3D ResNet architecture pretrained on video data transfer effectively to the task of diagnosing pulmonary abnormalities in CT scans?

Key findings

• Cov3d achieved a macro F1 score of 0.9476 on the validation set for detecting the presence of COVID-19, significantly outperforming the MIA-COV19D 2022 competition baseline.
• For severity classification, the model achieved a macro F1 score of 0.7552, which exceeds the baseline of 0.63 from the MIA-COV19D 2022 competition.
• The highest-performing model for detection used a resolution of 160×320×320 and a loss weighting λ=0.1 favoring the detection task.
• The best severity classification model used a resolution of 128×256×256 with λ=0.9, indicating that greater emphasis on severity loss improved performance on that task.
• Ensemble inference using the four highest-performing models for each task was planned for test set submission, suggesting improved robustness through model averaging.
• The model demonstrated strong generalization, with consistent performance across multiple hyperparameter configurations, indicating stability in training and inference.