[Paper Review] AI-MIA: COVID-19 Detection & Severity Analysis through Medical Imaging

Dimitrios Kollias, Αναστάσιος Αρσένος+1 more|Lincoln Repository (University of Lincoln)|Jun 9, 2022

COVID-19 diagnosis using AI4 citations

TL;DR

This paper presents a CNN-RNN deep learning baseline for COVID-19 detection and severity classification using a new large-scale 3D CT scan database (COV19-CT-DB) of ~7,700 scans. The model processes sequential CT slices via ResNet50 for feature extraction and a GRU for temporal modeling, achieving 0.77 macro F1 on detection and 0.63 on severity classification, establishing a benchmark for future AI-based medical imaging research in COVID-19.

ABSTRACT

This paper presents the baseline approach for the organized 2nd Covid-19 Competition, occurring in the framework of the AIMIA Workshop in the European Conference on Computer Vision (ECCV 2022). It presents the COV19-CT-DB database which is annotated for COVID-19 detction, consisting of about 7,700 3-D CT scans. Part of the database consisting of Covid-19 cases is further annotated in terms of four Covid-19 severity conditions. We have split the database and the latter part of it in training, validation and test datasets. The former two datasets are used for training and validation of machine learning models, while the latter will be used for evaluation of the developed models. The baseline approach consists of a deep learning approach, based on a CNN-RNN network and report its performance on the COVID19-CT-DB database.

Motivation & Objective

• To develop a robust deep learning baseline for automated detection of COVID-19 from 3D chest CT scans.
• To enable severity classification of COVID-19 cases using a newly curated, large-scale CT scan database.
• To provide a standardized benchmark dataset (COV19-CT-DB) with expert annotations for both detection and severity grading.
• To evaluate the performance of a CNN-RNN architecture on multi-label, multi-instance medical image classification tasks in a clinical context.
• To support the development of interpretable and generalizable AI models for radiological decision support in pandemic response.

Proposed method

• A 3D CT scan is processed as a sequence of 2D axial slices, with each slice resized to 224×224×3 and normalized to [0,1] after Hounsfield unit clipping (window/level: 350/−1150 HU).
• ResNet50 is used as the CNN backbone to extract spatial features from each 2D slice, followed by global average pooling, batch normalization, and dropout (keep probability 0.8).
• A one-directional GRU with 128 units processes the sequence of CNN-learned features across all slices to model temporal dependencies in the 3D volume.
• An output layer with two units and softmax activation produces class probabilities for COVID-19 vs. non-COVID-19 diagnosis.
• An 'at-least-one' voting scheme is applied: if any slice is predicted as COVID-19, the entire scan is classified as positive.
• The model is trained using Adam optimizer with a base learning rate of 10⁻⁴ and binary cross-entropy loss on a Tesla V100 GPU with a batch size of 5.

Experimental results

Research questions

• RQ1 Can a CNN-RNN architecture effectively classify 3D chest CT scans for COVID-19 presence using only whole-scan labels and no slice-level annotations?
• RQ2 How well does the proposed model generalize to unseen CT scans in terms of detection accuracy and severity classification?
• RQ3 What performance can be achieved using a large, multi-center, multi-scanner 3D CT database with expert-annotated severity levels?
• RQ4 To what extent does the sequential modeling of CT slices via RNN improve diagnostic performance over standard 2D or 3D CNNs?
• RQ5 How do different voting strategies (e.g., at-least-one vs. majority) impact final classification outcomes in a multi-instance learning setup?

Key findings

• The baseline CNN-RNN model achieved a macro F1 score of 0.77 on the validation set for the COVID-19 detection task, indicating strong performance on the binary classification of CT scans.
• For the more complex severity classification task, the model achieved a macro F1 score of 0.63, demonstrating feasibility but also room for improvement in multi-class severity prediction.
• The dataset COV19-CT-DB contains approximately 7,700 3D CT scans, with 258 in the training set and 61 in the validation set for detection, and similar splits for severity classification.
• The model's reliance on an 'at-least-one' voting scheme ensures that even a single suspicious slice can trigger a positive diagnosis, aligning with clinical sensitivity priorities.
• The use of ResNet50 as a feature extractor combined with GRU-based sequential modeling enables effective 3D context understanding despite the absence of slice-level annotations.
• The study establishes a reproducible baseline for future research in AI-driven medical imaging for infectious lung diseases, particularly in the context of pandemic response.