CovTANet: A Hybrid Tri-Level Attention-Based Network for Lesion Segmentation, Diagnosis, and Severity Prediction of COVID-19 Chest CT Scans

Highlights

• A single end-to-end network jointly handling all three clinically-relevant tasks: lesion segmentation, diagnosis (normal vs. COVID), and severity prediction (mild vs. severe) from chest CT volumes.
• A novel tri-level attention unit (TAU) combining channel, spatial, and pixel attention, applied repeatedly through both the segmentation network and the joint diagnosis/severity classifiers.
• A two-phase optimization scheme: a tri-level attention segmentation network (TA-SegNet) is trained first, then frozen and integrated into a joint framework that uses its lesion features to guide diagnosis and severity prediction — letting the classifiers focus on infected regions instead of the whole CT volume.
• TA-SegNet reaches 62.3% Dice on MosMedData (11.8% over U-Net, 26.7% over FCN) and 90.2% Dice on a secondary dataset; the full CovTANet reaches 85.2% diagnostic accuracy even on mild-symptom cases (95.8% for severe cases) and 91.7% accuracy distinguishing mild vs. severe severity.

Results

Evaluated on MosMedData (1110 CT volumes from Moscow hospitals; 50 volumes with pixel-level lesion annotations; volumes further labeled normal/mild/severe) with 5-fold cross-validation, plus a secondary 20-volume public segmentation dataset for additional validation.

Segmentation Ablation Study

Starting from a plain U-Net baseline (V1), each proposed component is added incrementally:

Encoder and decoder TAUs each help individually (+4.1% and +2.9% Dice respectively) and more together (+6.6%); fusing multiscale encoder–decoder features adds another +4.4%, combining with two-stage TAUs for +9.7%; and an EfficientNet backbone adds a further +2.1% — each component contributing independently to the final result.

Segmentation vs. Other Networks

TA-SegNet's 99.6% sensitivity means it very rarely misses a lesion outright, while also holding the highest precision of any method tested — most competing networks trade one for the other, producing either excess false positives or missed lesions.

Qualitative Segmentation

Fig. 8 — Qualitative segmentation comparison. Green = true positive, blue = false positive, red = false negative. TA-SegNet ("PROPOSED") shows visibly fewer false positives/negatives than FCN, U-Net++, U-Net, Inf-Net, and MultiResUNet across both datasets.

Segmentation on a Secondary Dataset

Joint Diagnosis & Severity Prediction

Diagnosis is reported separately for mild and severe cases to highlight early-diagnosis performance, since distinguishing mild COVID from normal CT is the harder and more clinically valuable case:

Every baseline here operates directly on raw CT slices; CovTANet's advantage — roughly 20–25 points of accuracy over the best baseline in every column — comes specifically from feeding in TA-SegNet's lesion-focused features instead, letting the classifier ignore the large fraction of each CT volume that is healthy tissue and irrelevant to the diagnosis.

Citation

@article{mahmud2021covtanet,
  title   = {{CovTANet}: A Hybrid Tri-Level Attention-Based Network
             for Lesion Segmentation, Diagnosis, and Severity
             Prediction of {COVID-19} Chest {CT} Scans},
  author  = {Mahmud, Tanvir and Alam, Md. Jahin and Chowdhury, Sakib
             and Ali, Shams Nafisa and Rahman, Md. Maisoon and
             Fattah, Shaikh Anowarul and Saquib, Mohammad},
  journal = {IEEE Transactions on Industrial Informatics},
  volume  = {17},
  number  = {9},
  pages   = {6489--6498},
  year    = {2021},
  doi     = {10.1109/TII.2020.3048391}
}