Predicting Treatment Response to Image-Guided Therapies Using Machine Learning: An Example for Trans-Arterial Treatment of Hepatocellular Carcinoma
October 10th, 2018
Intra-arterial therapies are the standard of care for patients with hepatocellular carcinoma who cannot undergo surgical resection. A method for predicting response to these therapies is proposed. The technique uses pre-procedural clinical, demographic, and imaging information to train machine learning models capable of predicting response prior to treatment.
This method can help predict response to intra arterial therapies. Machine learning applications in interventional oncology will change the way we treat liver cancer and other diseases with image-guided therapies. This method has the potential to improve the way two important decisions are made.
Generally, individuals new to this method will struggle because of the need to generate the features and to perform machine learning and code. Visual demonstration of this method is critical as the feature generation steps entail extracting those features using imaging masks. A well-annotated and clinically representative dataset is necessary to train the model.
More patient data and large cohorts will be necessary to further improve such a machine learning algorithm. After deciding which clinical features to include in the model, use an appropriate natural language tool kit to parse the plain-text clinic notes into sentences to allow searching for terms of interest. Then store each patient's features in a file with one feature per line.
For non-binary features, obtain the median value of each feature across all of the patients, and binarize each feature as a true or false value based on the median value. To calculate a mean liver enhancement feature from a medical image. After isolating the voxels containing the liver, enter the mean liver enhancement command.
To determine the liver volume feature, enter the commands as indicated. Then calculate median values for each imaging feature and binarize the features as demonstrated. For aggregation and reduction of the features, first remove the low-variance features from consideration and read the features in the binary matrix.
Next, use the variance threshold model to compute the features appearing in at least 20%of both responders and non-responders, similar to that illustrated in the table. Remove any features with a low univariate association with the outcome and filter only those features that pass the original low variant screening retaining n features where n is the number of patients. Then read in the binary matrix for each feature, similar to as illustrated in the table.
After the features have been binarized and filtered, it's time to train the model. Model training is a straightforward process that fits the features to the outcome under study and can be used to make predictions about new patients. In this representative analysis of 36 patients who underwent trans arterial therapies for hepatocellular carcinoma, 25 features were identified and binarized as demonstrated with 5 features satisfying both the variants and the univariate association filters.
Each patient was labeled as either a responder or non-responder under the quantitative European Association for the study of the liver response criteria. As illustrated, both the logistic regression and random forest models predicted a trans arterial chemoembolization treatment response with an overall accuracy of 78%Following this method, it is possible to make predictions on new patients by applying the training model. This machine learning technique can be universally applied to any therapy that involves pre-procedural, inter-procedural, and post-procedural imaging.
View the full transcript and gain access to thousands of scientific videos
Overview
This article discusses a method for predicting response to intra-arterial therapies in patients with hepatocellular carcinoma. By utilizing pre-procedural clinical, demographic, and imaging data, machine learning models can be trained to forecast treatment outcomes.
Key Study Components
Area of Science
- Interventional oncology
- Machine learning applications
- Hepatocellular carcinoma treatment
Background
- Intra-arterial therapies are standard for patients unable to undergo surgery.
- Predicting treatment response can enhance clinical decision-making.
- Feature generation from imaging data is crucial for model training.
- A clinically representative dataset is necessary for effective model performance.
Purpose of Study
- To improve prediction of patient response to intra-arterial therapies.
- To facilitate better clinical decision-making in liver cancer treatment.
- To demonstrate the importance of machine learning in interventional oncology.
Methods Used
- Pre-procedural data collection (clinical, demographic, imaging).
- Machine learning model training for response prediction.
- Feature extraction using imaging masks.
- Utilization of a well-annotated dataset for model accuracy.
Main Results
- The proposed method shows promise in predicting treatment outcomes.
- Machine learning can transform the approach to liver cancer therapies.
- Visual demonstrations are essential for understanding feature generation.
- Challenges exist for newcomers in implementing this method.
Conclusions
- This method has the potential to enhance treatment strategies for liver cancer.
- Machine learning applications can significantly impact interventional oncology.
- Further research and development are needed to refine these predictive models.
Frequently Asked Questions
Executive Industry Relevance
Predicting treatment response prior to intervention addresses a critical challenge in oncology drug development: reducing late-stage clinical failure due to lack of efficacy. By enabling data-driven patient stratification and go/no-go decisions, this machine learning framework supports mechanistic de-risking and improves predictive confidence in early discovery. Its applicability to image-guided therapies extends relevance across oncology modalities, enhancing portfolio-level risk-adjusted prioritization.
Strategic Applications in Biopharma R&D
Early Discovery & Target Validation
- Scientific Value: Supports therapeutic hypothesis interrogation by linking baseline features to treatment response, clarifying which patient subgroups are most likely to benefit.
- Operational Value: Enables functional target validation through retrospective analysis of clinical and imaging outcomes, reducing mechanistic ambiguity in early-stage programs.
- Predictive Value: Improves portfolio triage by identifying non-responders early, allowing resource reallocation to higher-likelihood indications.
Screening & Assay Development
- Scientific Value: Prepares validated biological and clinical systems for downstream workflows by establishing reproducible feature-response relationships.
- Operational Value: Promotes assay standardization and scalability through binarized feature extraction and cross-validation protocols.
- Screening Readiness: Generates quantitative outputs (e.g., prediction accuracy, feature importance) that enable reliable compound or intervention evaluation in preclinical models.
Translational & Preclinical Research
- Translational Continuity: Connects discovery-phase feature identification to preclinical validation by using clinically derived biomarkers to inform model training.
- Risk-Adjusted Advancement: Supports go/no-go decisions by quantifying prediction confidence, reducing biological risk before costly preclinical studies.
- Mechanistic De-risking: Focuses on predictive value rather than mechanism alone, aligning with translational biomarker strategies in oncology pipelines.
Pipeline & Workflow Integration
The method fits within the discovery continuum from target validation through lead identification to preclinical efficacy testing, particularly for oncology modalities involving image-guided delivery.
- Discovery Biology: Supports hypothesis testing by identifying which baseline clinical and imaging features correlate with treatment response, clarifying patient selection criteria.
- Screening: Enables assay readiness through standardized feature extraction and binarization, improving reproducibility across sites and models.
- Analytics: Delivers quantitative outputs such as prediction accuracy (78% in the study) and feature selection metrics, enabling objective comparison of intervention conditions.
- Translational Research: Connects to preclinical continuity by using response labels derived from established criteria (e.g., EASL) to align model outputs with clinically relevant endpoints.
- Enterprise Reuse: Frames the feature engineering and model averaging approach as a reusable capability applicable across indications and imaging-guided therapies beyond hepatocellular carcinoma.
Operational & Enterprise Impact
- Scientific Value: Increases predictive confidence in target validation by reducing reliance on anecdotal response data and enabling data-driven patient stratification.
- Operational Value: Enhances reproducibility and scalability through standardized feature extraction, binarization, and leave-one-out cross-validation.
- Strategic Value: Improves go/no-go decision quality, leading to better capital efficiency and reduced late-stage biological risk in oncology portfolios.
- Portfolio Impact: Enables risk-adjusted prioritization by identifying indications or subgroups with higher predicted response likelihood, supporting smarter advancement decisions.
Implementation Considerations
- Requires expertise in clinical data curation, feature engineering, and machine learning model training and validation.
- Depends on access to well-annotated, clinically representative datasets with paired imaging, demographic, and outcome data.
- Necessitates cross-team standardization in feature definition and binarization thresholds to ensure model generalizability.
- Involves adaptation considerations when applying the framework to different model systems or imaging modalities.
- Limited by the manual effort required to derive features from imaging and clinical notes, which may constrain scalability without automation.
