Deep learning applications in osteosarcoma MRI: A systematic review of recent advances in AI-based osteosarcoma diagnosis.
Authors
Affiliations (3)
Affiliations (3)
- School of Computer Sciences, Universiti Sains Malaysia, USM Penang, Malaysia.
- School of Electrical and Electronic Engineering, Engineering Campus, USM Nibong Tebal, Penang, Malaysia.
- Department of Radiology, School of Medical Sciences, Universiti Sains Malaysia, Kubang Kerian, Kelantan, Malaysia.
Abstract
Primary malignant bone tumours of the skeleton have a great diversity in their biological behaviour, and the most common in adolescence is osteosarcoma for which the diagnosis and therapeutic management are both a challenge. The use of machine and deep learning in image analysis for MRI, diffusion-weighted imaging (DWI) and dynamic contrast-enhanced MRI (DCE-MRI) to diagnose osteosarcoma has shown major improvements recently. This systematic review is performed according to PRISMA 2020 guidelines, and it comprises a critical analysis of 38 peer-reviewed articles from 2020 to 2025, which address the various aspects of tumour segmentation, classification, and prognosis. The methodological mapping revealed seven main categories that included numerous examples of radiomics (n = 8) and convolutional neural networks (CNNs, n = 6), as well as vision transformers and multimodal architectures for emerging technologies (n = 5 each). A quantitative evaluation of the segmentation performance showed that Multimodal achieved the best spatial overlap with an average Dice Similarity Coefficient (DSC) of 94.7% (range: 94.5-94.9%), while CNN of standard size was comparable, Lightweight networks had a DSC between 91.4 and 99.4% with an average of 91.9% and Vision Transformers a range of 90.5-94.9% with an average of 92.1%. By contrast, the performance gap among the Attention Mechanisms was greatest with a mean of 87.67% (64.6-98.5%), while the lowest was the mean 85.7% (81.6-89.8%) for Unsupervised Clustering. In addition to segmentation, the clinical nomograms based on radiomics also showed excellent prognostic potential:The AUC values for predicting pathological necrosis after chemotherapy varied from 0.807 to 0.848. More importantly it was found that there were significant translational gaps because all 38 models were limited to retrospective, single centre cohorts, with evaluation of risk of bias. Advanced algorithms perform well with localized precision but face a number of challenges with robust multicenter testing and a public repository.