The potential for different computed tomography-based machine learning networks to automatically segment and...

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Pelvic and sacral osteosarcomas (OS) and Ewing's sarcomas (ES) are rare bone cancers.
• Accurate differentiation between OS and ES is crucial for effective treatment planning.
• Computed tomography (CT) imaging is a primary modality for evaluating these tumors.

Purpose of the Study:

• To evaluate and compare machine learning (ML) and deep learning (DL) models for identifying pelvic and sacral OS and ES using CT data.
• To develop and validate an optimal deep learning model for accurate tumor classification.
• To assess the performance of AI models against expert radiologist diagnoses.

Main Methods:

• Analysis of CT scans from 185 patients with pathologically confirmed pelvic and sacral OS and ES.
• Comparison of 9 radiomics-based ML models, 1 radiomics-based convolutional neural network (CNN), and 1 3D CNN model.
• Development and evaluation of a 2-step nnU-Net model for automatic tumor segmentation and identification.

Main Results:

• The nnU-Net model achieved the highest performance, with an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.835 and an Accuracy (ACC) of 0.830.
• The radiomics-based CNN model (AUC=0.812, ACC=0.774) outperformed the 3D CNN model (AUC=0.709, ACC=0.717) and traditional ML models.
• The nnU-Net model's diagnostic accuracy was significantly higher than that of primary physicians.

Conclusions:

• The proposed nnU-Net model demonstrates superior performance in differentiating pelvic and sacral OS and ES.
• This deep learning approach offers an end-to-end, non-invasive, and accurate auxiliary diagnostic tool.
• AI-driven analysis of CT scans holds significant potential for improving the diagnosis of rare bone sarcomas.