Skip to content

Deep learning in medical imaging: multimodal segmentation and assisted diagnosis

    Recent HRIA research, conducted in partnership with medical universities, aims to improve diagnostic accuracy through neural architectures optimized for various types of scans (CT, MRI, X-ray, OCT).

    Vascular Network Segmentation and Radiographic Analysis

    Structural improvements are introduced to standard models to increase sensitivity in detecting fine anatomical structures:

    Angio-OCT (Retinal Networks): The I-MedSAM model was used, augmented with a Frangi module. This modification led to a +1% improvement in Dice, Precision, and Sensitivity metrics for vascular network segmentation.

    Chest Radiographs: An improved YOLO_V12 architecture was developed for detecting thoracic pathologies (cardiomegaly, aortic dilation, pulmonary nodules, and pleural effusion).

    Scoliosis Assessment: Use of YOLO OBB (Oriented Bounding Boxes) combined with a modified I-MedSAM model for automatic Cobb angle calculation.

    Tumor Recognition through Domain Adaptation Learning (DAL)

    A major contribution consists in the use of domain adaptation learning techniques for liver and pancreatic tumor recognition:

    • Dataset: Contrast-enhanced CT images from patients with liver tumors (45 patients) and pancreatic tumors (30 patients).
    • Methodology: A mixed dataset was used, followed by fine-tuning through DAL.
    • Performance: Models such as ResNet50, ResNet101, InceptionV3, and EfficientNetB0 were evaluated comparatively. The results indicate that the DAL approach (Acc_DAL) consistently outperforms the accuracy of models trained in a standard way or on mixed datasets.
    Liver tumors – maximum accuracy of 89.2%
    Pancreatic tumors – maximum accuracy of 98.9%

    Research Directions and Assisted Diagnosis (UMF Partnership)

    In collaboration with UMF Cluj, AI-assisted diagnostic systems are being developed for complex liver pathologies:

    • Focal Lesions: Classification of hepatic lesions using abbreviated MRI protocols.
    • Hepatic Fibrosis: Classification of fibrosis stage through multimodal analysis (CT/MRI/Ultrasound).
    • Treatment Monitoring: Evaluation of treatment response for hepatic lesions using CT scans.
    • Portal Hypertension: Automated diagnosis of portal hypertension using CT data.