EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Deep Learning Framework for Breast Tumor Detection and Segmentation Using CT Imaging

Zainab Pervaiz, Tallat Jabeen, Ruqaiyya Adil, Muhammad Waleed Khan, Aaysha Tariq, Muhammad Irfan Zafar, Ishrat Jabeen

IEEE Access · 2026 · 14 · 79138-79154

DOI: 10.1109/access.2026.3691800 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Zainab Pervaiz und 6 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 71 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Key Research and Development Program of China, National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Breast cancer remains the leading malignancy among women globally, yet AI-based detection systems focus predominantly on mammography while whole-body CT imaging is routinely performed for oncological staging thus, remains critically underexplored. Existing CT-based studies are limited by small datasets, geographical bias, and lack of spatial localization capabilities. This study introduces a radiologist-verified CT scan dataset of 9330 annotated slices from three Pakistani hospitals, approved by institutional ethics committees. Data was partitioned into training (70%), validation (15%), and test (15%) sets to develop a two-stage deep learning framework. The training data includes 6,106 whole-body CT slices (3,053 tumor, 3,053 normal). The perfectly balanced training dataset eliminates class-level imbalance, while detection-level imbalance is addressed through distributional focal loss mechanisms. We systematically evaluated seven transfer learning architectures (VGG16, VGG19, ResNet50V2, InceptionV3, DenseNet121, DenseNet201, EfficientNetB0–B5), revealing that VGG16 achieved optimal performance (82% accuracy, 0.80 AUC-ROC) compared to deeper architectures (ResNet50V2: 49%, EfficientNet: 50%), demonstrating that simpler models exhibit better inductive bias for constrained medical imaging datasets. Building upon this, we developed an integrated YOLOv11-Seg framework exploiting architectural innovations: C3K2 bottleneck modules for hierarchical feature extraction, Spatial Pyramid Pooling Fast (SPPF) for multi-scale context aggregation, and Pixel Spatial Attention (C2PSA) for enhanced boundary delineation. YOLOv11-Seg achieved 84.6% precision, 90% recall, 0.84 mAP@50, 0.839 Dice coefficient, and 0.724 mean IoU. Error analysis revealed false negatives (3.83%) primarily from low-contrast lesions and small tumors ( 10,000 frames demonstrated 10 FPS processing, 93.1% sensitivity, 97.9% specificity, and 94.3% temporal consistency, enabling clinical workflow integration. Limitations of the current study include single-radiologist annotation, geographical homogeneity, retrospective design, and binary classification scope, warranting prospective multi-center validation, multi-class classification extension, and explainable AI development. This work establishes foundation for AI-assisted opportunistic breast cancer screening in whole-body CT imaging, particularly valuable for resource-constrained settings, with clear pathways toward comprehensive staging and diagnostic support systems.

Förderung

  • National Key Research and Development Program of China
  • National Natural Science Foundation of China

Zitieren

APA

Pervaiz, Z., Jabeen, T., Adil, R., Khan, M. W., Tariq, A., Zafar, M. I., & Jabeen, I. (2026). Deep Learning Framework for Breast Tumor Detection and Segmentation Using CT Imaging. IEEE Access, 14, 79138-79154. https://doi.org/10.1109/access.2026.3691800

Vancouver

Pervaiz Z, Jabeen T, Adil R, Khan MW, Tariq A, Zafar MI, et al. Deep Learning Framework for Breast Tumor Detection and Segmentation Using CT Imaging. IEEE Access. 2026;14:79138-79154. doi:10.1109/access.2026.3691800

BibTeX
@article{Pervaiz2026,
  author = {Pervaiz, Zainab and Jabeen, Tallat and Adil, Ruqaiyya and Khan, Muhammad Waleed and Tariq, Aaysha and Zafar, Muhammad Irfan and Jabeen, Ishrat},
  title = {Deep Learning Framework for Breast Tumor Detection and Segmentation Using CT Imaging},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {79138--79154},
  doi = {10.1109/access.2026.3691800}
}
Kostenlos starten →