CovidCTNet: an open-source deep learning approach to diagnose covid-19 using small cohort of CT images

The study addresses the urgent need for accurate, rapid Covid-19 diagnosis beyond RT-PCR, which has a 30–35% error rate and cannot distinguish contamination from active disease. CT imaging can detect lung involvement and disease stage but suffers from limited specificity due to similarities between Covid-19 and other viral pneumonias such as CAP. The authors hypothesize that machine learning can mitigate technical bias and human error in CT interpretation, enabling reliable differential diagnosis. They propose CovidCTNet, a framework designed to assist radiologists through (i) accurate, rapid detection, (ii) compatibility with diverse CT hardware and imaging protocols, and (iii) usability without specialized computer science expertise. Leveraging BCDU-Net to operate effectively with small, heterogeneous datasets, the approach aims to improve diagnostic performance over conventional radiologist reads and prior deep learning methods that require large datasets.

Prior work has established higher sensitivity of chest CT than RT-PCR in Covid-19 diagnosis but with limited specificity due to overlapping imaging features (e.g., ground-glass opacities, consolidation) between Covid-19 and CAP. Multiple AI approaches using CT have been proposed, often relying on large datasets and various segmentation strategies (e.g., U-Net, DRUNET, FCN, SegNet, DeepLabv3) and classifiers (e.g., ResNet-18, ResNet-50, EfficientNet, 3D ResNet-18). Reported performances vary (AUC ~87–97%, sensitivities ~84–92%, specificities ~82–96%). The literature highlights the need for generalizable methods that handle heterogeneous imaging from multiple centers and minimize bias, with transparent, reproducible code and adherence to medical imaging AI reporting guidelines. This work positions CovidCTNet as an open-source, small-cohort-capable alternative emphasizing robust preprocessing to improve specificity.

Study design and data: Under IRB approval (IR.TUMS.VCR.REC.1399.007) with informed consent, CT scans were collected from five Iranian medical centers and the publicly available SPIE-AAPM-NCI LUNGx dataset. The dataset comprised 335 patients and 16,750 CT slices: Covid-19 (n=111; 5,550 slices; RT-PCR and radiologist confirmed), CAP/other viral (n=115; 5,750 slices), and Controls (n=109; 5,450 slices). An additional 70 LUNGx scans were included; 66 (21,888 slices) were used for training/validation of the preprocessing model and 4 for reader tests. Preprocessing pipeline: (1) Resampling to isotropic 1×1×1 mm voxel spacing to harmonize diverse scanner protocols. (2) Intensity windowing to Hounsfield Units [−1000, 400] to keep lung-relevant ranges (air ≈ −1000 HU, lung ≈ −700 to −600 HU; exclusion of non-lung tissues), followed by min–max normalization to [0,1]. (3) 2D slice resizing to 128×128 pixels for x–y dimensions; the z-dimension (number of slices) was initially preserved. BCDU-Net training with pseudo-infection (Perlin noise): To enable infection extraction despite limited labeled data, a BCDU-Net (Bi-directional ConvLSTM U-Net with densely connected convolutions) was trained to reconstruct clean control lungs and cancel noise. Perlin noise was injected into a subset of control slices to simulate pseudo-infection. Inputs to BCDU-Net were mixed noisy and non-noisy control images; targets were the corresponding original noise-free control images. The model learned to (a) remove non-lung tissues (e.g., heart, skin, table) and (b) suppress pseudo-infection noise, effectively learning normal lung structure. Training used 66 control patients (21,888 slices); the trained BCDU-Net was then frozen. Infection extraction by subtraction: After applying the same preprocessing to all patients (Covid-19, CAP, Control), each CT slice was passed through the frozen BCDU-Net to obtain de-noised, reconstructed lung slices. The difference between the preprocessed input slices and BCDU-Net outputs highlighted infection regions (Covid-19 or CAP) while suppressing non-lung tissues and artifacts. Slice-wise infection maps were concatenated along z to form 3D volumes. 3D CNN classification: Due to variable number of slices per study, volumes were resized to a uniform shape of 50×128×128 for the 3D CNN input. Labels were radiologist-assigned (Covid-19, CAP, Control). The CNN architecture consisted of multiple 3D convolutional and max-pooling layers followed by dense layers with dropout. Training used a 95%/5% train/validation split (hold-out). A balanced training subset used 100 cases per class (Control, CAP, Covid-19; total 300) with validation of 5 per class (total 15). A separate reader test set of 20 cases (4 Control, 10 CAP, 6 Covid-19) was used for human–AI comparison. Optimization details: Adam optimizer with learning rate 0.001 was used during preprocessing model training (as reported). All code, model parameters, and usage instructions are open-source at https://github.com/mohofar/covidctnet.

- Preprocessing necessity: Incorporating BCDU-Net and Perlin noise markedly improved validation performance on small data. With BCDU-Net, two-class accuracy was 82% versus 58% without; precision/recall/F1 for Covid-19 also improved (example using 50 cases/class). - Validation (algorithmic hold-out): When classifying Covid-19 vs non-Covid-19 (two classes), AUC was 94% with accuracy 93.33%, sensitivity 90.91%, specificity 100%. For three-class classification (Covid-19 vs CAP vs Control), accuracy was 86.66%. - Reader study (n=20 cases): CovidCTNet outperformed four independent radiologists. Two-class (Covid-19 vs non-Covid-19) accuracy was 95% for CovidCTNet vs 81% average radiologist accuracy; three-class accuracy was 85% for CovidCTNet vs 71% radiologist average. Model sensitivity/specificity for Covid-19 were 93%/100% vs radiologists’ 79%/82.14%. - ROC (reader test, three classes): AUCs were Control 0.97, Covid-19 0.90, CAP 0.98; macro-average ≈0.96. - The framework generalized across heterogeneous CT scanners (12 models from five brands) and operated effectively on small cohorts (335 patients total).

CovidCTNet addresses the challenge of distinguishing Covid-19 from CAP and other lung diseases on CT, where overlapping features (e.g., GGOs, consolidation) reduce specificity. By training BCDU-Net to reconstruct normal lung structure using pseudo-infection (Perlin noise) on control images, the method highlights true infection regions through subtraction, enabling a downstream 3D CNN to focus on relevant pathology. This strategy reduces overfitting on small datasets and mitigates scanner and protocol heterogeneity. The model achieved high validation performance and outperformed expert radiologists in a blinded reader study, demonstrating clinical utility as a decision-support tool. Open-source release promotes transparency, reproducibility, and rapid community-driven improvement, aligning with medical AI reporting guidelines. The findings suggest that robust preprocessing and lesion-focused representations can substantially enhance CT-based Covid-19 detection accuracy and reliability in diverse clinical settings.

The study introduces CovidCTNet, an open-source, multi-stage deep learning pipeline that leverages BCDU-Net-based cleaning and noise cancellation with Perlin noise augmentation to extract infection regions, followed by 3D CNN classification. The approach delivers high accuracy and AUC for Covid-19 detection, robust performance on small, heterogeneous datasets, and superior results to radiologists in a reader study. CovidCTNet’s design supports broad deployment across varied CT hardware and facilitates community enhancement. Future work will expand and diversify the dataset beyond predominantly Iranian cases, integrate patient demographics (age, gender, medical history) for predictive modeling, and validate across larger multi-center CT databases to further assess generalizability.

- Dataset composition is relatively small and predominantly from Iranian patients, potentially limiting generalizability; broader multi-center, international validation is needed. - The infection extraction relies on Perlin noise as a proxy for anomalies during BCDU-Net training, which may not capture the full spectrum of real pathologies. - Raw CT data are not publicly available due to privacy constraints, limiting external replication beyond the provided sample images; access requires a data use agreement. - Variable CT acquisition parameters were harmonized via preprocessing, but residual domain shifts may persist. - Reader study size (n=20) is limited; larger prospective evaluations are warranted.