Automated Morphological Structure Detection, Segmentation, and 3D Model Reconstruction from Medical Images Using Artificial Intelligence Deep Neural Network Technologies. Doctoral Thesis

Abstract

This research investigates contemporary approaches of automated medical image analysis and the creation of accurate anatomical models. The study aimed to develop and validate artificial intelligence-based methodologies for the automated detection, segmentation, and quantification of diverse morphological structures across various imaging modalities (computed tomography, magnetic resonance tomography, and histological slides). Another significant aim was the creation of extensive protocols for the reconstruction of three-dimensional (3D) anatomical models derived from radiological imaging data. The research results, including 3D printed models and software tools, were intended to improve medical education and clinical diagnostics. This research has established several protocols for reconstructing 3D anatomical models from data acquired via diverse imaging modalities, including computed tomography, micro- computed tomography, structured-light 3D scanning, and photogrammetry. These digital models were validated through 3D printing and their effective application in anatomy education. Deep neural network architectures, such as U-Net, and YOLO, were systematically trained and evaluated. These models were applied to tasks that included the segmentation of vertebrae and spinal metastatic lesions from computed tomography scans, and the precise quantification of interstitial cells in histological slides. The developed methodologies were further integrated into software tools to demonstrate practical utility and facilitate broader application. Key findings demonstrate the efficacy of AI models. In vertebral segmentation, the U-Net architecture achieved an F-beta score of 0.96 – meaning that it correctly identified 96 % of vertebral structures while keeping false positives to a minimum – whereas for lytic metastasis detection it reached 0.68, indicating more modest sensitivity and precision in spotting lesions. In histopathology, YOLO-based detectors attained a mean average precision at a 50 % overlap threshold (mAP₅₀) of 92 %, signifying that 92 % of cellular structures were both correctly detected and localised with at least half-area agreement. Finally, the 3D reconstruction workflows consistently produced printed models whose anatomy matched the original specimens to educationally useful degree. This research results confirms that advanced image processing techniques that incorporate deep neural networks can provide reliable, accurate, and reproducible results in the detection and segmentation of morphological structures. The methodologies and software developed offer significant potential to improve medical image analysis, anatomical education, and advance clinical diagnostics.