Platinum-based chemotherapy remains the standard first-line treatment for advanced lung cancer patients who lack targetable mutations, with cisplatin and carboplatin forming the backbone of most regimens. These drugs can produce meaningful initial responses.
However, platinum resistance is a major clinical problem - many patients either do not respond from the start or develop resistance during treatment. Clinically, resistance is defined as disease progression or recurrence within 6 months of completing chemotherapy.
Early identification of patients likely to develop resistance is critically important but currently lacking in routine practice. Physicians currently cannot tell before starting treatment which patients will respond and which will not, meaning patients who will ultimately fail platinum chemotherapy often miss the window to receive alternative treatments earlier.
This study investigated whether features extracted from pre-treatment CT scans using radiomics - combined with standard blood test results - could identify distinct patient subtypes with different rates of platinum resistance, enabling more targeted treatment planning.
Radiomics is a field that uses computer algorithms to extract large numbers of quantitative features from medical images - such as CT scans - that are invisible to the human eye. These include measurements of texture, shape, and intensity patterns within and around a tumor.
Unlike a radiologist's interpretation, radiomics can analyze hundreds or thousands of image characteristics simultaneously, detecting subtle patterns that may correlate with underlying tumor biology, gene expression, or treatment outcomes.
In this study, 1,228 radiomic features were extracted from each patient's pre-treatment CT scan, spanning histogram features (intensity distributions), morphological features (shape characteristics), texture features (spatial patterns), and wavelet features (multi-scale structural characteristics).
Wavelet features are a particularly powerful class - they analyze the image at multiple resolutions simultaneously, capturing complex spatial patterns that reflect tumor heterogeneity more comprehensively than simpler feature sets. The majority of features in this study were wavelet-based.
This retrospective cohort study analyzed 684 lung cancer patients treated at Xiangya Hospital of Central South University in China between January 2011 and June 2025. The cohort was 78% male, with a median age of 56 years. Cancer types included adenocarcinoma (50%), squamous cell carcinoma (44%), and other types (5%). Critically, 98% of patients had stage III or IV disease.
Platinum resistance status was the primary outcome of interest: 320 patients (47%) were classified as platinum-resistant (disease progression or recurrence within 6 months of completing chemotherapy) and 364 (53%) as platinum-sensitive.
An extraordinarily comprehensive set of blood tests was collected for each patient before treatment: 39 biochemical indicators, 9 coagulation function tests, 3 tumor markers, and 22 complete blood count parameters - 73 laboratory values in total. This allowed a systematic search for blood biomarkers of platinum resistance.
For CT image analysis, two experienced radiologists independently manually delineated the 3D tumor boundary on each patient's pre-treatment scan. Regions of hemorrhage, necrosis, and cystic degeneration within the tumor were excluded to focus the analysis on viable tumor tissue. Measurements were normalized to account for differences between the three CT scanner types used over the 14-year study period.
Hierarchical clustering of the 1,228 radiomic features was used to group patients based on their CT imaging patterns. The optimal number of clusters was determined mathematically using the silhouette coefficient, which measures how well each patient fits within its assigned cluster versus neighboring clusters.
The analysis consistently identified two distinct imaging-based subtypes, confirmed by principal component analysis showing clear separation of the two groups in image feature space. Every patient was assigned to one of these two groups based on their tumor's imaging characteristics alone.
Predictive models were then built using four approaches: univariate analysis (individual blood tests), multivariate analysis (statistically selected blood tests), LASSO regression (a machine learning method for feature selection that balances predictive power against model complexity), and an all-variable model (incorporating all 73 laboratory measurements).
Additional combined models integrated radiomic features with blood tests, and further versions incorporated the patient's subtype category as an additional predictor. This systematic comparison allowed the researchers to determine the independent contribution of each data type to platinum resistance prediction.
Clustering revealed two subtypes with significantly different platinum resistance rates and clinical laboratory profiles. The first was named the 'imaging-physiological homeostasis subtype' - characterized by lower creatine kinase, myoglobin, and CEA levels, higher CO2 levels, and a platinum resistance rate of approximately 41%.
The second was called the 'imaging-physiological disequilibrium subtype,' characterized by relatively higher creatinine, creatine kinase, myoglobin, and CEA levels, lower CO2 levels, and a significantly higher platinum resistance rate of approximately 51%. The lower CO2 suggests a tendency toward metabolic acidosis, while higher creatinine suggests reduced renal reserve capacity.
These biological differences make physiological sense: patients in the disequilibrium subtype show markers suggesting greater tissue damage (elevated muscle enzymes), higher tumor burden (elevated CEA), and metabolic stress (lower CO2/acidosis tendency). These systemic signs of physiological strain may reflect a more aggressive tumor biology that is more likely to resist treatment.
Interestingly, the platinum resistance rate differences were consistent regardless of whether patients received cisplatin alone, carboplatin alone, or both drugs in combination - suggesting the subtype reflects intrinsic tumor biology rather than drug-specific resistance mechanisms.
Comparing prediction approaches across the whole patient cohort, radiomics-only models achieved an AUC (predictive accuracy measure) of 0.578 on the test set - only slightly better than chance. Clinical laboratory models alone performed substantially better at AUC 0.805. The combined model integrating both data types reached AUC 0.811.
Within each subtype separately, the combined radiomics plus laboratory model achieved AUC values of 0.821 (homeostasis subtype) and 0.822 (disequilibrium subtype) - both better than models built on the full unsegmented population. This demonstrates the value of analyzing subtypes separately rather than pooling all patients together.
Adding the subtype label as an additional predictor in models built on all patients provided incremental improvements across all model combinations, suggesting that the imaging subtype captures independent biological information about platinum resistance beyond what blood tests and radiomic features capture individually.
Decision curve analysis - which measures whether using a model to guide clinical decisions produces a net benefit for patients over standard practice - showed that all four blood test models outperformed the conventional approach of simply classifying patients as 'platinum-sensitive' or 'platinum-resistant' based on clinical judgment alone.
The predictors of platinum resistance differed between subtypes, highlighting the importance of subtype-specific analysis. In the homeostasis subtype, blood glucose and anion gap were the independent predictors in multivariate analysis - suggesting that metabolic dysregulation (elevated glucose and acid-base imbalance) predicts resistance within this relatively physiologically stable group.
In the disequilibrium subtype, the independent predictors were calcium, beta-2 microglobulin, and alpha-fetoprotein. Beta-2 microglobulin reflects cellular turnover and can indicate rapid tumor growth, while abnormal calcium may reflect bone involvement or parathyroid hormone-related protein production by the tumor.
These subtype-specific differences suggest that platinum resistance is driven by different biological mechanisms in different patient groups - a finding consistent with the known complexity of platinum resistance, which is understood to arise through multiple distinct pathways.
Clinically, this implies that a single universal biomarker for platinum resistance may not exist, and that the best predictive approach must account for the patient's underlying tumor biology - which is now accessible through CT imaging analysis before treatment begins.
The study's main limitations include its single-center design, which means the model has not been validated on patients from other hospitals or healthcare systems. External validation is essential before the approach could be considered for routine clinical use.
For patients with multiple tumor lesions, only the largest lesion was analyzed, which may not fully capture the overall heterogeneity of a patient's disease. Metastatic sites may have different imaging characteristics and resistance patterns than the primary tumor.
The study did not analyze genomic, proteomic, or metabolomic data, which are known to be important in platinum resistance. Integrating these molecular profiles with CT radiomics and blood tests in future work could further improve predictive accuracy.
Future research should conduct multi-center external validation, integrate multi-omic molecular data, and link imaging subtypes to progression-free survival outcomes - building toward a clinically actionable tool that could guide treatment selection before therapy begins.