Mobile-Friendly Solution for COVID-19 Detection from Computed Tomography Images
14th Computing Conference, CC 2026, London, İngiltere, 9 - 10 Temmuz 2026, cilt.1951 LNNS, ss.19-31, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Cilt numarası: 1951 LNNS
- Doi Numarası: 10.1007/978-3-032-24810-7_2
- Basıldığı Şehir: London
- Basıldığı Ülke: İngiltere
- Sayfa Sayıları: ss.19-31
- Anahtar Kelimeler: Computed tomography images, COVID-19 diagnosis, Macro F1 Score, xxs_Mobile ViT transformer
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
In this study, we propose a mobile-friendly, modern, and accurate solution for COVID-19 diagnosis using Computed Tomography (CT) images. In our proposed solution, we leveraged state-of-the-art Transformer models in computer vision, namely “mobile ViT xxs Transformer”, to reach our goal. Two key steps of image processing were employed to decrease model misclassifications. Firstly, the uppermost and lowermost slices of each CT scan were removed, preserving sixty per cent in each scan. Secondly, all slices underwent rectangular manual cropping to bring focus on the lung areas or Region of Interest (ROI) of the slices. Subsequently, resized CT scans (384 by 384) were input into the transformer. The transformer model is suited for grayscale input images and a binary classification task. To determine the overall diagnosis for each patient, majority voting was deployed for each CT scan to make predictions for each patient. To verify the efficiency of our method, a big and rigorously annotated database of CT images, named COV19-CT-DB, was used. Verification was made on both the validation partition and the test partition of unseen images of the database. The model’s performance exceeded the baseline on the given dataset. These results present our solution as accurate and suitable for clinical and personal usage. The code can be found at https://github.com/IDU-CVLab/COV19D_4th.