Pipeline for converting medical tomographic data for custom implant design

Abstract

This paper presents a comprehensive methodology for the automated conversion of computed tomography (CT) series (DICOM format) into monolithic 3D polygonal models (STL). The system is designed to facilitate the direct additive manufacturing of patient-specific bone implants and the creation of verified training datasets for neural network-based spatial reconstruction algorithms. The issue of tissue optical density discrepancies between native and contrast-enhanced diagnostic phases is investigated. An algorithmic solution is proposed to eliminate the “contour breach” effect in epiphyseal joint zones characterized by thin cortical bone - a phenomenon that, in conventional approaches, leads to hollow defects or the distortion of the joint space. A multi-level algorithm is introduced, encompassing spatial coordinate standardization, tissue hysteresis filtering based on the core-shell ratio, and multi-axial filling of internal cavities within isolated anatomical components. The use of the standardized NIfTI format is justified as an integration bridge connecting the initial CT data processing module, deep learning models, and the final polygonal mesh generator. A predictive module based on machine learning (ML) algorithms has been developed; it automatically generates the optimal prosthesis geometry, performs topological optimization of its internal porosity, and selects a biocompatible material using a multi-criteria approach, all based on extracted morphometric vectors, local Hounsfield Unit (HU) optical density distributions, and biomechanical load tensors. To overcome the “black box” effect, an Explainable AI (XAI) subsystem - based on feature importance metrics (SHAP/LIME values) - has been implemented to generate a detailed analytical report. This report fully elucidates the cause-and-effect relationships between the patient's anatomical parameters and the model's final decision, justifying the selection of the porosity gradient, structural element thickness, and alloy physicochemical properties for clinical verification. The architecture of an autonomous software system featuring hardware-accelerated graphics for clinical application is described.