Using Smart Imaging Analysis to Predict Chemotherapy Response in Bone Cancer: A Patient's Guide to Radiomics for Osteosarcoma

Table of Contents

• Key Points
• Understanding Osteosarcoma and Its Treatment
• The Challenge of Predicting Chemotherapy Response
• What Is Radiomics and How Does It Work?
• How This Research Review Was Conducted
• Key Findings: What the Studies Show
• MRI Shows Superior Capability
• Clinical Implications: What This Means for Patients
• Study Limitations: What This Review Couldn't Prove
• Recommendations for Patients and Families
• Frequently Asked Questions
• Source Information

Key Points

• MRI-based radiomics predicted neoadjuvant chemotherapy response in osteosarcoma with up to 93.8% accuracy and AUC 0.961 in one study.
• Radiomics extracts computer-analyzed patterns from CT, MRI, or PET scans that may predict ≥90% tumor necrosis before surgery.
• Response to chemotherapy is currently assessed after surgery; radiomics aims to predict it earlier, allowing treatment personalization.
• Most radiomics studies had small, single-center samples, and the approach is not yet standardized for routine clinical use.
• PET/CT radiomics with machine learning improved prediction, but raw PET features alone performed poorly in one study.

Understanding Osteosarcoma and Its Treatment

Osteosarcoma is the most common primary malignant bone tumor, accounting for approximately 20% of all primary malignant bone tumors. It primarily affects children and adolescents, with the metaphysis (the growing end) of long bones—such as the thigh bone or shin bone—being the typical site where the cancer develops.

Modern treatment strategies have significantly improved outcomes. The current 5-year survival rate for osteosarcoma patients is now 60%–70%, a major improvement from past decades. This progress is largely due to the standard treatment approach, which combines three components:

• Neoadjuvant chemotherapy (NAC) — chemotherapy given before surgery to shrink the tumor
• Surgical resection — surgically removing the tumor
• Postoperative adjuvant chemotherapy — additional chemotherapy given after surgery to eliminate any remaining cancer cells

The current clinical standard uses multi-agent chemotherapy regimens, typically including high-dose methotrexate (HD-MTX), ifosfamide (IFO), doxorubicin (ADM), and cisplatin (DDP). These drugs work together to kill cancer cells, but the treatment is intense and can cause significant side effects.

Here is where the problem lies: patients respond very differently to the same chemotherapy regimen. The response rate—defined as the percentage of patients achieving at least 90% tumor cell death (necrosis) after chemotherapy—ranges from just 30% to 60%. This means a substantial number of patients endure the harsh side effects of chemotherapy without getting the full benefit.

The Challenge of Predicting Chemotherapy Response

Chemotherapy sensitivity is one of the most critical prognostic factors in osteosarcoma. The landmark EURAMOS-1 clinical trial found that only about 50% of patients achieve a good histological response, defined as 90% or more tumor necrosis after neoadjuvant chemotherapy. This response rate has important implications for prognosis, as it guided treatment decisions in the trial.

Currently, the standard method for assessing chemotherapy response relies on examining the surgically removed tumor under a microscope. This approach, described by Huvos and colleagues, classifies patients into two groups:

• Good pathological responders — patients with more than 90% tumor necrosis
• Poor pathological responders — patients with less than 90% tumor necrosis

The fundamental limitation of this approach is timing. The response can only be assessed after surgery, which means it cannot guide decisions about whether to continue or change chemotherapy during treatment. To date, no definitive clinical, biological, or imaging markers allow doctors to reliably predict chemotherapy response early enough to modify treatment strategies.

It's also important to note that "response to chemotherapy" encompasses two distinct concepts in osteosarcoma:

• Histological response — defined by tumor cell death measured after surgery
• Radiological response — reflected by imaging changes, such as alterations in tumor size or density on CT or MRI scans

Radiomics-based predictive tools primarily aim to predict the histological response using pre-treatment imaging. This bridges the gap between early radiological assessment and delayed surgical pathology results.

Existing radiological criteria such as RECIST (Response Evaluation Criteria in Solid Tumors) are less applicable to osteosarcoma because the presence of bone matrix complicates the assessment of tumor size changes. This makes it difficult to tell whether a tumor is responding simply by measuring its dimensions on a scan.

What Is Radiomics and How Does It Work?

Radiomics represents a revolutionary shift in medical imaging analysis. Instead of relying on a radiologist's visual interpretation alone, radiomics uses advanced computer algorithms to extract thousands of quantitative measurements from medical images. These measurements fall into several categories:

• Morphological features — shape characteristics of the tumor
• Intensity features (first-order statistics) — statistical properties of pixel values within the tumor
• Texture features (second- and higher-order features) — patterns of pixel arrangements that reveal tumor heterogeneity

These features can be extracted from multiple imaging modalities, including Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and 18F-Fluorodeoxyglucose Positron Emission Tomography (18F-FDG PET/CT). The technology can precisely characterize tumor heterogeneity and the tumor microenvironment, overcoming the subjectivity and reproducibility limitations of traditional manual image interpretation.

To understand how radiomics works, it helps to know about the artificial intelligence (AI) tools behind it:

• Machine Learning (ML) — a subset of AI that uses algorithms to learn patterns from data and make predictions without being explicitly programmed for each task
• Deep Learning (DL) — a specialized branch of ML that uses multi-layer neural networks to automatically extract hierarchical features from complex data like medical images
• Convolutional Neural Networks (CNNs) — a type of deep learning algorithm that uses multi-layer convolutional kernels to automatically extract image features. This approach enables simultaneous analysis of both microscopic texture characteristics and macroscopic morphological patterns

Conventional statistical methods have significant limitations when processing the high-dimensional, nonlinear data characteristic of radiomics. Their capacity for feature extraction is restricted, and they struggle to recognize complex patterns. In contrast, deep learning approaches offer two distinct advantages:

1. They eliminate the selection bias inherent in manual feature engineering
2. Through backpropagation optimization, they can directly identify potential imaging biomarkers from raw pixel data

Clinical validation studies have established that deep learning-based prediction models demonstrate superior performance in assessing neoadjuvant chemotherapy response compared to traditional methods.

How This Research Review Was Conducted

The researchers followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines to ensure their review was thorough and unbiased. Here's how they conducted their search:

• Databases searched: Three major medical databases
• Time period covered: Articles published between January 1, 2014, and December 31, 2024
• Search terms used: "osteosarcoma AND chemotherapy AND response AND prediction"
• Language restriction: English-language original research articles only; review articles were excluded

Two independent investigators performed the initial screening based on titles and abstracts. Potentially eligible studies then underwent full-text review. Any discrepancies between the reviewers were resolved through consensus discussion.

Studies were included if they met all three criteria:

1. Histologically confirmed primary osteosarcoma
2. Evaluation of neoadjuvant chemotherapy response prediction
3. Reported pathological response (e.g., ≥90% tumor necrosis)

Studies were excluded if they:

• Did not report relevant outcomes
• Did not specifically include the term "osteosarcoma"
• Limited assessment to post-chemotherapy imaging only
• Were prognostic rather than predictive studies
• Used non-imaging-based approaches

For data extraction, two investigators independently collected information using a standardized form. This included publication year, study design, patient demographics, imaging parameters, radiomic feature extraction methodology, predictive model construction, and primary outcomes. When data was missing or ambiguous, they attempted to contact the original authors for clarification.

The researchers also used the Radiomics Quality Score (RQS) tool to evaluate study quality. The assessment focused on several critical components, including imaging acquisition protocols (equipment parameters and scanning standardization), tumor segmentation methods, feature extraction techniques, and model building approaches.

Finally, the authors performed a quantitative meta-analysis using random-effects models to compare the predictive performance of different imaging modalities (MRI and PET-CT). They calculated pooled sensitivity, specificity, and area under the receiver operating characteristic curve (AUC)—a measure of how well a test distinguishes between good and poor responders. Study heterogeneity (inconsistency between studies) was assessed using I² statistics.

Key Findings: What the Studies Show

The review identified 24 eligible studies evaluating radiomics for predicting neoadjuvant chemotherapy response in osteosarcoma. The studies used various imaging modalities, including X-ray, CT, MRI, nuclear medicine scans, and PET/CT. Here are the key findings organized by imaging type.

X-ray and MRI Combined

A 2023 study by Z. Luo and colleagues involving 102 patients combined X-ray and MRI radiomics with clinical data. The model using all data combined achieved an AUC of 0.828. Individual models performed as follows: clinical plus X-ray radiomics reached an AUC of 0.760, X-ray radiomics alone 0.751, MRI radiomics alone 0.706, and the combination of X-ray plus MRI radiomics 0.796.

CT-Based Radiomics

F. Yang and colleagues (2024) conducted a two-center study with 225 patients using CT radiomics. Models combining radiomics and clinical features achieved AUC values of 0.78 in the training set and 0.75 in the independent validation set.

D. Fu and colleagues (2023) studied 18 patients and examined peri-osteosarcoma fat characteristics on CT. The peri-osteosarcoma fat attenuation index (FAI) achieved an AUC of 0.950, while the 6-hour methotrexate concentration had an AUC of 0.963—both indicating excellent predictive performance for chemotherapy response.

L. Xu and colleagues (2021) studied 157 patients and improved prediction accuracy by combining CT radiomic features from both the tumor area and surrounding non-tumor bone areas, using multiple machine-learning techniques.

Nuclear Medicine Imaging (Tc-MIBI)

C. Wu and colleagues (2019) studied 30 patients using Tc-MIBI scans. They found that the tumor washout rate (how quickly the radioactive tracer leaves the tumor) was negatively correlated with tumor necrosis rate (r = −0.510, P = 0.004). When a washout rate of 25% or less was used as the threshold for predicting good chemotherapy response, the sensitivity was 100%, specificity was 91.7%, and accuracy was 95.8%.

MRI-Based Radiomics

Multiple MRI-based studies demonstrated strong predictive performance:

• F. Zheng (2024), 106 patients: A deep learning radiomics model achieved the highest prediction performance with an accuracy of 93.8% and an AUC of 0.961 in the test set.
• Y. Zhang (2024), 109 patients: A combined model using pre- and post-chemotherapy MRI achieved AUC values of 0.999 in the training set and 0.915 in the test set. Notably, post-chemotherapy models performed better than pre-chemotherapy models.
• Kanthawang Thanat (2024), 95 patients: Tumor volume greater than 150 mL and maximum axial diameter greater than 7.0 cm were identified as independent predictors of poor response (P = 0.025 and P = 0.045, respectively).
• J. Zhong (2022), 144 patients: Radiomics models achieved AUC values of 0.699, 0.759, and 0.784 across different model configurations.
• K.Y. Teo (2022), 15 patients: Machine learning combined with multimodal MRI showed excellent performance in predicting tumor necrosis, with AUC of 0.999 in the training set and 0.915 in the test set.
• Esha Baidya Kayal (2022), 35 patients (IVIM-MRI): Baseline parameters showed an AUC of 0.87, sensitivity of 86%, and specificity of 77%. After the first chemotherapy cycle, the AUC increased to 0.96, with sensitivity of 86% and specificity of 100%.
• J. Dufau (2019), 69 patients: Of these patients, 55.1% (38 out of 69) were good histological responders. A support vector machine model based on initial MRI radiomic data achieved an AUROC of 0.98, sensitivity of 100% (95% CI [100%–100%]), and specificity of 86% (95% CI [59.7%–111%]).
• G.J. Djuričić (2017), 22 patients: Computational morphological analysis using fractal and gray-level co-occurrence matrix algorithms predicted chemotherapy response with an AUC of 0.82 and accuracy of 82%.

Advanced MRI Techniques (DCE-MRI and IVIM-DWI)

• Zeng Yan-Ni (2022), 25 patients (DCE-MRI): Using thresholds of 3.2%/s (Slope), 175 seconds (TTP), and 5.4% (ER), the sensitivity and specificity for predicting good chemotherapy response were 83.3% and 92.3%, 91.7% and 69.2%, and 84.6% and 75.0%, respectively.
• Xibin Xia (2022), 163 patients (DCE-MRI + IVIM-DWI): After two treatment cycles, Ktrans, Kep, and Ve values were significantly lower in the complete response/partial response group than in the stable disease/progressive disease group. Meanwhile, D, ADC, and f values were significantly higher in good responders. Blood tests for alkaline phosphatase (ALP) and lactate dehydrogenase (LDH) were positively correlated with Ktrans, Kep, and Ve values, but negatively correlated with D, ADC, and f values.
• A. Bouhamama (2022), 176 patients across 3 centers: The combined DCE-MRI and IVIM-DWI model achieved AUC values of 0.95 and 0.97, with sensitivity of 91% and specificity of 92%.
• L. Zhang (2021), 102 patients (DCE-MRI): Models combining clinical risk factors (such as surgical stage) with radiomic features outperformed radiomics alone. The combined model achieved prediction accuracy of 0.91 in the training set and 0.90 in the test set, with AUC values of 0.94 and 0.95, respectively.
• M.M. Saleh (2020), 53 patients (multiparametric MRI with DWI): This study highlighted the importance of using multi-parameter MRI, especially diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) maps, in predicting chemotherapy sensitivity. It also pointed out the limitations of traditional assessment methods based solely on tumor volume change.

PET/CT-Based Radiomics

• B.C. Kim (2021), 52 patients: Chemotherapy response test accuracy using image texture features was 0.83, and metastasis test accuracy was 0.76. Combining imaging with genomic data (radiogenomics) improved results: the highest test accuracy for chemotherapy response with AUC_max, KI67, and EZRIN was 0.85 with an AUC of 0.89. Metastasis prediction accuracy increased by 10% using radiogenomics data.
• Kim J (2021), 105 patients: While a prediction model based on baseline PET texture features with traditional machine learning estimated poor outcomes, a 2D CNN network using baseline FDG PET images could predict treatment response before chemotherapy. This suggests deep learning on PET images can help decide whether to perform neoadjuvant chemotherapy.
• H. Song (2019), 35 patients: Metabolic tumor volume (MTV) was identified as the best parameter for predicting chemotherapy response, with an AUC of 0.918 and a p-value of less than 0.0001, indicating very high predictive accuracy.
• S.Y. Jeong (2019), 70 patients: Baseline PET features alone (SUVmax, TLG, MTV, first-order entropy, and gray-level co-occurrence matrix entropy) had poor AUCs ranging from 0.536 to 0.553. However, machine learning features using linear SVM, random forest, and gradient boost achieved AUCs of 0.72, 0.78, and 0.82, respectively—showing that machine learning dramatically improved predictive power.
• I. Lee B (2018), 62 patients: This study compared Tc-MDP bone scintigraphy and FDG PET/CT, showing that each imaging method (with features like T/N ratio and SUVmax) offered different advantages in predicting response.
• J.C. Davis (2018), 34 patients: SUVmax on routine images at 5 or 10 weeks, and the percentage change in SUVmax from baseline to week 10, were identified as metabolic predictors of histological response.
• B.H. Byun (2015), 34 patients (dual-phase FDG PET/CT): Using combined criteria of percentage SUV change and retention index (RI) values, prediction accuracies were 81%, 77%, and 77% for various combinations. The histological response after neoadjuvant chemotherapy could be predicted using RImean before treatment initiation, and the combined use of SUV and RI values provided better prediction.

MRI Shows Superior Capability

A key conclusion of this review is that MRI demonstrates superior capability in multimodal data integration and high-throughput feature extraction. This makes sense because MRI provides excellent soft tissue contrast, allowing detailed visualization of the tumor within the bone and surrounding tissues.

The strongest performing models in the entire review were primarily MRI-based. For instance, the deep learning radiomics model by F. Zheng achieved 93.8% accuracy and an AUC of 0.961, while Y. Zhang's combined pre- and post-chemotherapy MRI model reached an AUC of 0.999 in training and 0.915 in testing. The DCE-MRI study by A. Bouhamama across three centers achieved AUC values of 0.95 and 0.97.

PET/CT-based radiomics also showed promise, particularly when combined with machine learning or deep learning approaches. The study by S.Y. Jeong demonstrated that while raw PET features performed poorly (AUC 0.536–0.553), applying machine learning algorithms boosted performance dramatically (AUC 0.72–0.82). This highlights a crucial principle: it's not just the imaging modality that matters—the analytical approach is equally important.

Clinical Implications: What This Means for Patients

The clinical value of chemotherapy response prediction spans three key dimensions:

1. Treatment optimization. Accurate predictive models would enable truly personalized therapeutic strategies. For patients predicted to respond well to chemotherapy, treatment can be optimized—for example, maintaining effective regimens to avoid unnecessary escalation. For patients predicted to be poor responders, they could be promptly transitioned to alternative therapies, such as targeted agents or immunotherapy, which are currently being investigated in clinical trials.

2. Prognostic management. Reliable prediction tools would facilitate early identification of high-risk patient populations, allowing for more intensive monitoring and timely clinical interventions. For example, the SARC024 trial has used predictive tools to stratify patients into novel therapy arms based on predicted resistance, accelerating the drug development process.

3. Research translation. Robust prediction systems provide valuable stratification criteria for clinical trial design. This means future clinical trials can be designed more efficiently, with the right patients assigned to the right treatment arms from the start.

The noninvasive nature of radiomics is another major advantage. Unlike biopsy, which requires an invasive procedure, medical imaging can be repeated throughout therapy without additional risk to the patient. This allows for dynamic monitoring of how the tumor is responding and adjustment of treatment plans in real time.

Study Limitations: What This Review Couldn't Prove

While the results are encouraging, the review highlights several important limitations that patients should understand:

• Small sample sizes: Many of the included studies had small patient populations. For example, several studies had fewer than 35 patients, which limits the statistical power and generalizability of their findings.
• Single-center designs: The majority of studies were conducted at a single institution, which raises questions about whether the results would hold true across different hospitals with different imaging equipment and protocols.
• Heterogeneity across studies: The studies varied significantly in terms of imaging parameters, segmentation methods, feature extraction techniques, and machine learning algorithms. This heterogeneity complicates direct comparisons and meta-analysis.
• Lack of standardization: There is currently no uniform standard for how radiomic features should be extracted and analyzed, which can lead to reproducibility issues.
• Feature selection variability: The review noted that feature selection methods and validation strategies varied considerably across studies, affecting prediction accuracy.
• Distinction between predictive and prognostic: The review carefully distinguished between predictive studies (which forecast treatment response) and prognostic studies (which forecast overall outcomes). Only predictive studies were included, but this distinction can be difficult to maintain in practice.

It's also important to note that while radiomics shows excellent promise, these tools are not yet ready for routine clinical use. They should be considered investigational approaches that may complement—but not yet replace—standard histopathological evaluation after surgery.

Recommendations for Patients and Families

If you or a loved one is facing osteosarcoma treatment, here are some practical takeaways from this research:

1. Ask about imaging-based response assessment. While the standard of care remains histopathological evaluation after surgery, some cancer centers are beginning to use advanced imaging analysis to monitor response during neoadjuvant chemotherapy.
2. Participate in clinical trials when possible. Many of the machine learning and radiomics approaches discussed in this review are being refined in clinical trials. Participating in research can give you access to cutting-edge predictive tools while contributing to scientific progress.
3. Understand the response criteria. When your care team discusses "good response" to chemotherapy, they are typically referring to 90% or more tumor necrosis seen under the microscope after surgery. This is a key prognostic indicator.
4. Ask about MRI vs. PET/CT. Based on this review, MRI-based radiomics appears to offer the strongest predictive power, but PET/CT with machine learning analysis also shows promise. Your center's equipment and expertise will influence which approach is most feasible.
5. Discuss the limitations honestly. No imaging test or radiomics model can perfectly predict chemotherapy response. The goal of these tools is to provide additional information to guide treatment decisions, not to replace clinical judgment.
6. Consider seeking care at a specialized sarcoma center. Osteosarcoma is a rare cancer, and centers with high patient volumes are more likely to have access to advanced imaging analysis capabilities and clinical trials.

The bottom line: Radiomics represents an exciting frontier in osteosarcoma care. The ability to predict chemotherapy response from a simple MRI or PET scan—before any treatment is given—could spare patients from ineffective therapies, guide more personalized treatment plans, and ultimately improve survival outcomes. While more research is needed to standardize these approaches and validate them across larger, multi-center populations, the evidence to date is highly encouraging.

Frequently Asked Questions

What is radiomics and how can it help in osteosarcoma?

Radiomics uses computer algorithms to extract detailed information from medical images that the eye cannot see, such as texture patterns. In osteosarcoma, radiomics aims to predict whether chemotherapy given before surgery will achieve a good response, meaning at least 90% tumor cell death. This could help personalize treatment and avoid ineffective therapy.

Can radiomics predict response before surgery, while chemotherapy is still ongoing?

Yes. The goal is to predict the delayed surgical pathology result using imaging obtained before or during chemotherapy. One study found that MRI after the first chemotherapy cycle improved prediction compared to before treatment. This allows earlier adjustment of treatment, unlike the standard method that only assesses the tumor after it is surgically removed.

Is radiomics ready for routine clinical use in osteosarcoma?

No, not yet. The review states these tools should be considered investigational. They may complement but not replace the standard histopathological evaluation after surgery. Limitations include small sample sizes, single-center designs, and lack of standardization. It is important to discuss with your care team whether advanced imaging analysis is available and suitable.

What are the main limitations of the research on radiomics for osteosarcoma?

Many studies had small patient numbers, for example fewer than 35 patients in several. Most were from a single center, so results might not apply elsewhere. There was no uniform standard for extracting and analyzing radiomic features, causing reproducibility concerns. The review advises that no imaging tool yet replaces clinical judgment.

Should I ask my doctor about radiomics or advanced imaging analysis during treatment?

You may ask whether your center uses imaging-based response assessment during neoadjuvant chemotherapy. Some cancer centers are beginning to use advanced imaging analysis to monitor response. However, the standard of care remains examining the surgically removed tumor after surgery. Participate in clinical trials if possible, as these tools are being refined in research settings.

Should I get a second opinion about my osteosarcoma treatment plan before starting neoadjuvant chemotherapy?

For osteosarcoma, chemotherapy response is currently known only after surgery, when tumor cell death is checked under a microscope. Since roughly 30–60% of patients do not achieve a good response (90% or more tumor necrosis), a second opinion before starting neoadjuvant chemotherapy can be valuable. An expert may review your imaging and discuss whether investigational radiomics-based tools, which are not yet standard, could suggest alternative approaches such as targeted therapy or clinical trials. MRI-based analysis shows strong predictive potential in research. Diagnostic Detectives Network provides independent expert second opinions.

Source Information

Original article title: Radiomics for predicting sensitivity to neoadjuvant chemotherapy in osteosarcoma: current status and advances.

Authors: Zhang P, Yao W, Li Z, Fan Y, Du X, Wang B, Zhang F, Hou J, Su Q.

Affiliations: Henan Cancer Hospital / Affiliated Cancer Hospital of Zhengzhou University, Zhengzhou, China; Zhengzhou University, Zhengzhou, China; Zhoukou Central Hospital, Zhoukou, China

Journal: Oncology Reviews, Volume 19, Article 1633211

Publication date: October 17, 2025

DOI: 10.3389/or.2025.1633211

Study type: Systematic review with meta-analysis

This patient-friendly article is based on peer-reviewed research. The original article was published under the Creative Commons Attribution License (CC BY), which permits sharing and adaptation with appropriate credit to the original authors and source. This patient summary was created to make the research findings more accessible to patients, families, and caregivers. It is not intended as medical advice—please consult your oncology care team for guidance specific to your situation.