A New Integrated Medical-Image Processing System with High Clinical Applicability for Effective Dataset Preparation in ML-Based Diagnosis

Abstract

We have developed a new medical-image processing system with high clinical applicability for effective dataset preparation in computer-aided diagnosis (CAD) based on machine learning (ML) techniques. Despite the wide range of application of ML-based CAD, it remains a challenge to apply this technique to clinical diagnosis of specific diseases or evaluation of specific medical practices. This is due to the lack of an effective framework for preparing a sufficient number of datasets. Consequently, there is absence of a guideline or standard procedure for processes such as image acquisition, anonymization, annotation, preprocessing and feature extraction. To address this ongoing issue, we proposed a system that was designed to integrate the special functions such as incremental anonymization, annotation assistance, and hybrid process flow for preprocessing and feature extraction. The incremental anonymization aimed to enable batch processing after image acquisition, reducing the daily workload of medical specialists in their institutions. Cross annotation and error correction were also supported even outside the medical institution by cogeneration of annotation sheets with anonymized images and by its OCR-based data-collection process. A hybrid process flow combining a simple manual operation and complemental automation algorithm was adopted to accelerate preprocessing and feature extraction. The system prototype successfully streamlined the dataset preparation process, including the generation of 3D breast-mesh closures and their associated geometric features. This substantial enhancement in efficiency demonstrates the system’s high clinical applicability and its potential to significantly contribute to the field of breast reconstruction evaluation.