CNN-based automated classification of SPECT bone scan images

Abstract

Functional medicine imaging has been successfully applied to capture functional changes in pathological tissues of the body in recent years. SPECT nuclear medicine functional imaging has the potential to acquire information about areas of concern (e.g., lesions and organs) in a non-invasive manner, enabling semi-automated or automated decision-making for the purposes of disease diagnosis, treatment, evaluation, and prediction. To reliably identify that whether or not at least one hotspot or lesion presents in a whole-body SPECT image, in this work, we develop a group of CNN-based classifiers. Specifically, we first propose a preprocessing method that transforms each original SPECT file into the required form by deep learning model, including normalization, 3-channel construction, rotation and scaling, size standardization, and size adapting. Second, six different classifiers are constructed by fine-tuning parameters of the standard VGG-16 model. Last, a group of real-world SPECT whole-body bone scan files were utilized to evaluate the developed classifiers. Experiment results shows that our classifiers are workable for the 2-class classification of SPECT images, achieving a best value of 0.7641, 0.6678, 1.000, and 0.6574 for defined evaluation metrics Acc, Pre, Rec, and AUC, respectively.