Comprehensive diagnostic model for osteosarcoma classification using CT imaging features

Abstract

Objective: The main objective of this study was to create and assess a detailed diagnostic model with an optimizing feature selection algorithm that combines computed tomography (CT) imaging characteristics, demographic information, and genetic markers to enhance the accuracy of benign and malignant classification of osteosarcoma. This research seeks to enhance the early identification and categorization of benign and malignant of osteosarcoma, ultimately enabling more personalized and efficient treatment approaches.

Methods: Data from 225 patients diagnosed with osteosarcoma at two different medical institutions between June 2018 and June 2021 were gathered for this research study. A novel feature selection approach that combined Principal Component Analysis (PCA) with Improved Particle Swarm Optimization (IPSO) was utilized to analyze 1743 image-derived features. The performance of the resulting model was evaluated using metrics such as area under the receiver operating characteristic curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE), and compared to models developed using conventional feature selection methods.

Results: The proposed model showed promising predictive performance with an AUC of 0.87, accuracy of 0.80, sensitivity of 0.75, and specificity of 0.85. These results suggest improved predictive ability compared to models built using traditional feature selection techniques, particularly in terms of accuracy and specificity. However, there is room for improvement in enhancing sensitivity.

Conclusion: Our study introduces a novel predictive model for distinguishing between benign and malignant osteosarcoma., emphasizing its potential significance in clinical practice. Through the utilization of CT imaging features, our model shows improved accuracy and specificity, marking progress in the early detection and classification of osteosarcoma as either benign or malignant. Future investigations will concentrate on enhancing the model's sensitivity and validating its effectiveness on a larger dataset, aiming to boost its clinical relevance and support personalized treatment approaches for osteosarcoma.

Keywords: Computed tomography (CT); Machine learning; Osteosarcoma; Particle swarm optimizer (PSO).

Fig. 1

Block diagram of our study.

Fig. 2

Different clusters with distribution of the top three features.

Fig. 3

Selected features and Their Weights.

Fig. 4

Confusion matrix for different feature extraction algorithm.

Fig. 5

Receiver Operating Characteristic (ROC) Curves.