A Stage IA and IB Endometrial Cancer Classification Method Using Genetic Programming on MRI Images.

BACKGROUND: Early diagnosis and accurate staging of endometrial cancer (EC) are crucial for effective treatment planning. Distinguishing stage IA from stage IB EC is challenging due to large variations in tumor and uterine morphology, as well as the limited availability of annotated magnetic resonance imaging (MRI) data for training robust models. METHODS: A genetic programming (GP)-based framework was developed for the classification of FIGO stage IA and IB EC using a small number of MRI images. The framework consists of three main components: (1) automatic detection of regions of interest (ROI) on MRI images, (2) GP-based feature extraction and construction, and (3) EC stage classification. A fast single-shot detector (SSD) was employed to automatically localize the uterus as the ROI. The detected ROI images were cropped and resized to construct training and test datasets. Four GP-based methods with different structures and primitives were employed for feature extraction and construction: GP with convolutional operators (COGP), GP with image descriptors (IDGP), GP with flexible program structures and image-related operators (FlexGP), and GP with automatic simultaneous learning of features and evolutionary ensembles (FELGP). The best-performing GP individuals were used to generate discriminative features, which were subsequently used to train classifiers for stage IA and IB EC. RESULTS: Experimental results from three MRI datasets revealed that GP-based methods achieved competitive performance relative to both neural and traditional non-neural machine learning approaches. The proposed methods achieved classification accuracies of up to 0.92 on cropped axial diffusion-weighted imaging (DWI), 0.87 on cropped axial T2-weighted imaging (T2WI), and 0.83 on cropped sagittal T2WI images of EC patients. CONCLUSIONS: GP-based methods effectively classify FIGO stage IA and IB endometrial cancer using limited MRI data. By automatically extracting discriminative and interpretable features from ROI within lesions, the proposed framework provides a reliable and transparent solution for EC staging, highlighting the potential of GP in medical image analysis and clinical decision support.