SOCH. An ML-based pipeline for PET automatic segmentation by heuristic algorithms means

Abstract

Nowadays, nuclear medicine procedures have become a standard for several pathologies, both for diagnosis and therapeutic purposes. Also, regarding therapeutic applications, the demand for novel techniques and new radioisotopes is increasing worldwide. Due to the high dose rates involved in therapy procedures, this aspect requires significant efforts related to the development of more accurate methods and protocols for individualized patient dosimetry estimations. New theranostic procedures allowing joint diagnosis/treatment implementation proves to be suitable for image-guided dosimetry. Therefore, appropriate image segmentation becomes a key issue for tissues/organs identification. Implementation of machine learning models for digital image processing is a promising opportunity to complement expert clinical analysis. This work presents SOCH, an original machine learning-based pipeline capable of PET/CT unsupervised automatic segmentation by heuristic algorithms means using clustering and machine learning techniques. Obtained results suggested, preliminary, that pipeline flows based on K-Means and HDBSCAN algorithms are capable of PET/CT image segmentation, proving to be a promising tool to assist expert clinicians in daily procedures.