Colorectal cancer remains a major public health challenge, and accurate quantification of T lymphocytes (CD3+,CD8+) within the tumor microenvironment holds important prognostic and predictive value. We present MS-HAF-Net, a multiscale hybrid dual-encoder architecture (a convolutional branch based on ResNet34 and a Swin Transformer branch) designed for both semantic segmentation and instance segmentation of lymphocytes in histopathological images. MS-HAF-Net fuses fine
local features with global contextual representations through multi-scale fusion, and is enhanced by a Residual Feature Augmentation (RFA) module and ECA attention blocks to improve detection of very small objects.
The network outputs multi-task heads (masks, contours, and center heatmaps) that enable robust instance separation via watershed post-processing. Comparative experiments against several baselines (YOLOv8-Seg, YOLOv11-Seg, SwinUNet, TransUNet, U-Net+EfficientNetB4) demonstrate that MS-HAF-Net reduces instance merging in dense regions while maintaining high localization accuracy. These results position MS-HAF-Net as a promising solution for integrating automated diagnostic assistance tools in oncology.
Index Terms—instance segmentation, histopathology, lymphocytes, CNN+Transformer, Swin, ResNet, MS-HAF-Net, ECA,RFA.

Auteurs : Feriel Lamirem , Wyssem Fathallah, Imen Helal, Nizar Sakli, Hedia Bellali, Hedi Sakli

Téléchargement : PDF

We conduct a comparative evaluation of chest X-ray and computed tomography (CT) scan imaging modalities for machine learning-based automatic COVID-19 diagnosis. Using a public corpus of 17,599 medical images, we applied transfer learning with MobileNetV2 to assess diagnostic accuracy, sensitivity, and specificity and established that CT scans are more accurate than X-rays, yet X-rays are more sensitive to non-COVID cases, and therefore have complementary roles in clinical practice.
Keywords— COVID-19 screening, Medical imaging, Chest X-ray, Computed tomography (CT) scan, Machine learning, Deep learning, Transfer learning, Image reprocessing, Artifact removal, Deblurring, AI-aided diagnosis, Data augmentation,Diagnostic accuracy, Multimodal imaging.

Auteurs : Jawhara Lazhar , Boutheina Ben-Ismail, Wyssem Fathallah, Hedi Sakli

Téléchargement : PDF