Delta-Radiomics Using Machine Learning Classifiers With Auxiliary Data Sets to Predict Disease Progression During Magnetic Resonance-Guided Radiotherapy in Adrenal Metastases

Abstract

PURPOSE

Adaptive radiotherapy accounts for interfractional anatomic changes. We hypothesize that changes in the gross tumor volumes identified during daily scans could be analyzed using delta-radiomics to predict disease progression events. We evaluated whether an auxiliary data set could improve prediction performance.

MATERIALS AND METHODS

We analyzed 108 patients (n = 90 internal; n = 18 external) who received ablative radiotherapy. The internal data set included 42 patients with adrenal cancer, 23 patients with lung cancer, and 25 patients with pancreatic cancer, with the clinical end point of progression-free survival events. The median dose was 50 Gy, which was delivered over five fractions. The delta features are the ratio of the features of the last to first treatment fraction, F5/F1, and the concatenation of the first and last fraction features, F1||F5. Decision tree classifier with and without auxiliary data sets, and the external data set was used exclusively for independent testing of the final models.

RESULTS

During internal training, for the F1||F5 model, the inclusion of the lung data set increased our AUC receiver operator characteristic curve (ROC) from 0.53 ± 0.12 to 0.61 ± 0.11, whereas the pancreatic data set increased our AUC-ROC to 0.60 ± 0.14. For the F5/F1 model, the inclusion of the lung auxiliary data increased our AUC-ROC from 0.52 ± 0.13 to 0.65 ± 0.11, whereas it modestly changed by 0.62 ± 0.13 with the pancreas. During external testing, for the F5/F1 model, we reported an AUC-ROC of 0.60 with the lung auxiliary data and 0.43 with the pancreatic data. Also, for the F5||F1 model, we reported an AUC-ROC of 0.70 with the lung auxiliary and 0.60 with the pancreatic data.

CONCLUSION

Decision trees provided an explainable model on the external data set. The validation of our model on an external data set may be the first step to biologically adapted radiotherapy recognizing radiomics signals for potential recurrence.

INTRODUCTION

The introduction of the magnetic resonance guided-linac (MRL) system has enabled high-precision tumor dose delivery while effectively sparing surrounding normal organs with superior soft tissue visualization compared with computed tomography (CT)–based radiotherapy platforms. Magnetic resonance–guided radiation therapy (MRgRT)¹ delivered on MRLs involves obtaining daily magnetic resonance images (MRIs) at each fraction, allowing real-time adjustments for tumor and organ positioning, thus enhancing treatment accuracy and reducing damage to surrounding healthy tissues.^2–6 Stereotactic MRI-guided adaptive radiation therapy (SMART) is an ultra-hypofractionated modality unique to MRLs that uses these advantages to deliver ablative doses accurately and precisely to tumors near radiosensitive structures.⁷ SMART has been proposed particularly for adrenal lesions because of their proximity to radiosensitive organs at risk with significant interfractional anatomic shifting (ie, gastrointestinal peristalsis of abdominal organs) and intrafractional motion management.^8–10

Radiomics involves extracting high-dimensional quantitative features from medical images, which provides insights into the pathology of cancer sites and potentially correlates with clinical outcomes like progression-free survival (PFS), overall survival, and so on.^11–15 Delta-radiomics can build upon this by incorporating changes in the radiomics features at different time points.¹⁶

We hypothesized that a machine learning (ML) model trained on an internal data set could classify patients on an external data set into groups with and without PFS events.

We, therefore, developed a delta-radiomics model using an interpretable decision tree ML classifier that can predict tumor progression events of metastatic adrenal tumors during daily MRgRT. We also deployed a multisite learning approach and evaluated the hypothesis of whether additional patients with lung or pancreatic cancer could be incorporated as auxiliary data sets during training to enrich the data set and overall prediction accuracy.

MATERIALS AND METHODS

Study Cohort

In total, 648 MRI scans from 108 patients who underwent SMART on a 0.35 T MRL (MRIdian, ViewRay Inc, Denver, CO) for either adrenal or pulmonary tumors from March 2019 to January 2023 were included for analysis. Six MRI scans were performed on each patient: one simulation scan and five therapeutic fraction scans (F1, F2, F3, F4, and F5). The simulation MRI was acquired 14–21 days before the first fraction, and then daily MRIs were acquired immediately before dose delivery at each fraction. Patient, tumor, and treatment characteristics are shown in Table 1.

TABLE 1.

Patients’ Information

Patients’ Characteristics	Internal Adrenal	Internal Lung	Internal Pancreas	External Adrenal
Cohort size, N	42	23	25	18
Age at treatment start, No. (range)	67 (19–82)	64 (32–79)	65 (60–73)	67 (28–79)
Gender, No. (%)
Female	20 (47.6)	13 (56.5)	12 (48)	8 (44.4)
Male	22 (52.4)	10 (43.5)	13 (52)	10 (55.6)
Tumor location lung, No. (%)
Right upper lobe		4 (17.4)
Right lower lobe		5(21.7)
Right medial lobe		1 (4.4)
Right hilum		2 (8.7)
Left upper lobe		5(21.7)
Left lower lobe		5(21.7)
Lingula		2 (8.7)
Mediastinum		1 (4.3)
Left lung, unknown		1 (4.3)
Pancreas, No. (%)
Head			21 (84)
Body			3 (12)
Tail			1 (4)
Adrenal, No. (%)
Right	16 (38.1)			8 (44.4)
Left	26 (61.9)			10 (55.6)
Primary tumor, No. (%)
NSCLC	14 (33.3)			14 (77.8)
SCLC	1 (2.4)			1 (5.6)
Bladder	1 (2.4)			1 (5.6)
Angiosarcoma				1 (5.6)
Breast	3 (7.1)
RCC	8 (19.0)
Melanoma	5 (11.9)
Prostate	1 (2.4)
NET	2 (4.8)
Merkel cell	2 (4.8)
ACC	2 (4.8)
Rectal adenocarcinoma	1 (2.4)
Pheochromocytoma	1 (2.4)
Skin SCC	1 (2.4)
Previous local therapy, No, (%)
Resection	7(16.7)	3 (13.0)
RT	3 (7.1)	4 (17.4)
Previous systemic therapy, No. (%)
Chemotherapy	31 (73.8)	16 (69.5)	25 (100)	11 (61.1)
Immunotherapy	26 (61.9)	3 (13.0)		8 (44.4)
Concurrent, No. (%)	13 (30.9)	3 (13.0)		5 (27.8)
RT dose, Gy
Total dose, median	50 (40–60)	50 (50–60)	50 (40–55)	50 (40–50)
BEDα/β=10, median	100 (72–132)	100 (100–132)		100 (72–100)
Target volumes, cm3
GTV, median	19.1 (2.1–105.6)	7.1 (1.5–20.8)		18.12 (7.0–48.6)
PTV, median	23.6 (3.5–135.0)	17.4 (4.3–38.7)		49.2 (15.4–81.6)
Clinical outcomes, months
Follow-up, median	14.1 (1.2–42.5)	16.7 (1.0–28.6)		6.8 (1.2–16.4)
Time to progression, median	8.15 (0.5–39.3)	5.00 (0.2–27.1)	6.97 (1.2–13.4)	4.96 (0.8–14.2)
Local failure, No. (%)	8 (19.0)	1 (4.3)		2(11.1)
PFS event, No. (%)	31 (73.8)	11 (47.8)	15 (60)	8 (44.4)

Abbreviations: ACC, adenoid cystic carcinoma; BED_α/β=10, biologically effective dose calculated with an α/β equal to 10; GTV, gross tumor volume; NET, neuroendocrine tumor; NSCLC, non–small-cell lung cancer; PFS, progression-free survival; PTV, planning tumor volume; RCC, renal cell carcinoma; RT, radiation therapy; SCC, squamous cell carcinoma; SCLC, small-cell lung cancer.

Clinical Outcomes

Patients were typically followed every 3–6 months with surveillance CT imaging. PFS events were measured from the end of RT and included locoregional disease progression, distant disease progression, and death. RECIST 1.1¹⁷ was used to evaluate response; however, any new distant lesion is also considered a progressing disease regardless of the size.

Delta-Radiomics Model

Two models were considered: (1) the concatenation of the first and last fractions, F1||F5, and (2) the ratio of the last and first fractions, F5/F1.

ML Classifier

This study classified patients into groups with and without PFS events using logistic regression, k-nearest neighbor, support vector machine, decision tree, and random forest classifier.

Survival Analysis and Data Augmentation

To overcome limited sample size and data imbalance, a novel approach was developed to train with a multisite data set, including auxiliary data sets and BorderlineSMOTE-1,¹⁸ respectively (Fig 1). The goal was to learn similar imaging characteristics and take advantage of the biological relationship between these sites. The study experimented with lung and pancreatic auxiliary sets, with lung data chosen due to the high frequency of lung metastasis to the adrenal gland, 80% of external patients with adrenal cancer who have lung cancer as the primary malignancy, and none of the patients in this study have primary pancreatic cancer. We used Kaplan-Meier survival analysis and log-rank P value to analyze discrepancies in time to event between the internal adrenal data set and external data set, as well as between the internal adrenal data set and internal auxiliary data sets.

FIG 1.

Machine learning model development workflow.

Model Development

The image preprocessing and feature extraction protocol can be found in the Data Supplement (Sections S1 and S2, respectively). This study adhered to TRIPOD type 3 criteria for model optimization and training, with an internal data set used for model optimization and training and an external data set used for independent and external testing. The external data set was tested following TRIPOD type 2a criteria.¹⁹ The 0.632+ bootstrapping method was used for feature selection, hyperparameter tuning, and model development. Our feature selection technique is Elastic Net Regression (Data Supplement, Section S3). Five-fold cross-validation was used for internal validation, and the AUC receiver operator characteristic curve (ROC) was used to compare and assess the performance of different delta-radiomics models. The consistency of the model’s external testing performance was validated when the testing AUC-ROC was within the standard deviation of the training results. The adrenal data set was split into 42 training and testing sets, with feature selection performed on the adrenal training set and hyperparameter tuning on the auxiliary data set. The ML model’s performance was strictly evaluated on the adrenal data set to avoid data leakage issues.

Data Collection Ethics

The Institutional Review Boards (IRBs) at the University of South Florida (IRB 20383), Moffitt Cancer Center (IRB 212790), and Miami Cancer Institute (IRB 1819730–3) approved and waived the informed consent requirement for the internal and external data set collection.

RESULTS

Patient, Tumor, and Treatment Characteristics

This analysis included 108 patients. Table 1 describes the baseline patient, tumor, and treatment characteristics. The internal data set consisted of 42 adrenal patients, the auxiliary data set consisted of 23 lung patients and 25 pancreatic patients, and the external data set consisted of 18 adrenal patients. There were eight external adrenal patients, 31 internal adrenal patients, 11 internal lung patients, and 15 internal pancreatic patients with disease progression events. Also, we have two external adrenal patients, eight internal adrenal patients, and one internal lung patient with local failure.

The log-rank test can explain the impact of the difference in the time-to-event in the data sets (Data Supplement, Section S4 and Figs S3-S5); there is no significant difference between the distribution of the internal and external adrenal data sets (P = .2064) and between internal adrenal and internal auxiliary pancreatic data (P = .6404). However, there is a significant difference between internal adrenal data and internal auxiliary lung data (P= .0029).

Internal Training

For optimal classifier selection, Table 2 shows the performance of the ML classifiers on the delta-radiomics F1||F5 model internal data set without and with the auxiliary data sets. Decision trees have a consistently small deviation from the mean AUC-AUC across the three data sets when compared with other classifiers. Therefore, the decision tree is our optimal classifier.

TABLE 2.

ML Classifier Selection—Performance Summary of ML Classifiers of F1||F5 Delta-Radiomics Model on Internal Adrenal Patients With and Without Auxiliary Data

ML Classifier	Adrenal Alone, AUC ± SD	Adrenal + Lung, AUC ± SD	Adrenal + Pancreas, AUC ± SD
Logistic regression	0.54 ± 0.21	0.62 ± 0.17	0.59 ± 0.17
Support vector classifier	0.65 ± 0.23	0.67 ± 0.12	0.58 ± 0.11
KNN	0.57 ± 0.18	0.68 ± 0.06	0.59 ± 0.11
Decision trees	0.53 ± 0.12	0.61 ± 0.11	0.60 ± 0.14
Random forest	0.52 ± 0.16	0.59 ± 0.14	0.57 ± 0.09

Abbreviations: KNN, K-nearest neighbor; ML, machine learning; SD, standard deviation.

Table 3 shows the results of the decision tree classifier on the internal adrenal data set with and without the auxiliary data set in the F1||F5 and F5/F1 delta-radiomics models. The AUC-ROC without auxiliary data was 0.53 ± 0.12 for the F1||F5 model, whereas it was 0.52 ± 0.13 for the F5/F1 model. The inclusion of the auxiliary data affects the prediction of the internal adrenal testing fold. For the F1||F5 model, with the lung data set, the AUC-ROC increased from 0.53 ± 0.12 to 0.61 ± 0.11, whereas the inclusion of the pancreatic data increased it to 0.60 ± 0.14. For the F5/F1 model, the inclusion of the lung auxiliary data increased our AUC-ROC from 0.52 ± 0.13 to 0.65 ± 0.11, whereas it was 0.63 ± 0.13 when the pancreatic data were included. In the F1||F5 and F5/F1 models, the evaluation of the internal adrenal test data set was highest when the lung data set was included as auxiliary data.

TABLE 3.

Internal Training/Optimal Model Selection—Performance Summary of Decision Tree Classifier on F1||F5 and F5/F1 Delta-Radiomics Models With and Without Auxiliary Data

Delta-Radiomics Dataset	ACC ± SD	AUC-ROC ± SD	NPV ± SD	PPV ± SD	f1-Score ± SD
F1||F5 delta-radiomics model
Adrenal BorderlineSMOTE-1	0.53 ± 0.06	0.53 ± 0.12	0.22 ± 0.13	0.78 ± 0.12	0.63 ±0.06
Adrenal + lung BorderlineSMOTE-1	0.67 ± 0.14	0.61 ± 0.11	0.48 ± 0.27	0.78 ± 0.09	0.76 ± 0.12
Adrenal + pancreatic BorderlineSMOTE-1	0.68 ± 0.14	0.61 ± 0.14	0.43 ± 0.33	0.79 ± 0.06	0.77 ± 0.11
F5/F1 delta-radiomics model
Adrenal BorderlineSMOTE-1	0.65 ± 0.12	0.52 ± 0.13	0.33 ± 0.37	0.75 ± 0.07	0.76 ± 0.09
Adrenal + lung BorderlineSMOTE-1	0.63 ± 0.12	0.65 ± 0.11	0.39 ± 0.09	0.86 ± 0.13	0.69 ± 0.14
Adrenal + pancreatic BorderlineSMOTE-1	0.56 ± 0.11	0.62 ± 0.13	0.35 ± 0.08	0.84 ± 0.14	0.63 ± 0.12

Abbreviations: ACC, accuracy; AUC-ROC, area under the receiver’s operating characteristic curve; NPV, negative predictive value; PPV, positive predictive value; SD, standard deviation.

External Testing

Testing our decision tree classification model on our external data set without the auxiliary data sets, we reported an AUC-ROC of 0.55 for the F1||F5 model and 0.57 for the F5/F1 model. With the inclusion of the auxiliary data, we observed an effect on the performance of our model.

For the F5/F1 model, we reported an AUC-ROC of 0.60 with the lung auxiliary data and 0.425 with the pancreatic data. Also, for the F5||F1 model, we reported an AUC-ROC of 0.70 with the lung auxiliary and 0.60 with the pancreatic data (Table 4).

TABLE 4.

External Testing—Performance Summary of Decision Tree Classifier on the External Adrenal Data

Delta-Radiomics Dataset	F1||F5					F5/F1
	ACC	AUC	NPV	PPV	f1-Score	ACC	AUC	NPV	PPV	f1-Score
Adrenal BorderlineSMOTE-1	0.56	0.55	0.50	0.60	0.60	0.61	0.57	0.25	0.90	0.72
Adrenal + lung BorderlineSMOTE-1	0.72	0.70	0.50	0.90	0.78	0.61	0.60	0.50	0.70	0.67
Adrenal + pancreatic BorderlineSMOTE-1	0.61	0.60	0.50	0.70	0.67	0.39	0.43	0.75	0.10	0.15

Abbreviations: ACC, accuracy; NPV, negative predictive value; PPV, positive predictive value.

We observed a better result with the lung auxiliary data. This may be considered as a method of improving the data harmonization between the internal and external data sets. We chose the F1||F5 model with the lung auxiliary data as our optimal model.

Model Interpretability

On external testing, the F1||F5 model has an AUC-ROC of 0.55 without auxiliary data, which was improved by the inclusion of the lung auxiliary data with an AUC-ROC of 0.70 and pancreatic auxiliary data with an AUC-ROC of 0.60. Therefore, our optimal delta-radiomics model is the F1||F5 model with the auxiliary lung data. The positive predictive value is 0.90, which implies that the probability of the model identifying a progression event is high and reliable, but the negative predictive value is 0.50, which implies that the model can only fairly classify a patient without progression. This may be due to the prevalence of the cases with progression events in the training population.

The top features used in the model prediction are a combination of features from the first and last fractions (Data Supplement, Fig S2). They are F1-Gray Level Small Zone Matrix (GLSZM) Small Zone Low Gray Level Emphasis, F1-Gray-Level Co-Occurrence Matrix (GLCM) avgCooccurence Joint Max, F5-GLSZM Zone Size Variance, F5-Minimum Histogram Gradient Gray Level, F1-Volume at Intensity Fraction 90, F1-Volume at Intensity Fraction Difference, and F5-GLCM Difference Variance. The GLSZM features describe the number of connected voxels that share the same intensity value; the GLCM features describe the joint probability occurrence of specified pixel pairs; and the intensity histogram features describe the distribution of discretized gray values.¹⁴ Figure 2 demonstrates a set of rules used by the ML decision tree classifier to classify the patients according to whether a progression event is present or not. Twelve of 18 test data set patients were correctly predicted through this approach.

FIG 2.

F1||F5 adrenal and auxiliary lung decision tree model–gives the splitting rules learned from the training data for prediction on the external data set. GLCM, Gray-Level Co-Occurrence Matrix; GLSZM, Gray-Level Small Zone Matrix; PFS, progression-free survival.

DISCUSSION

MRgRT has thus far focused on the management of interfractional anatomic shifts and intrafractional tumor motion to personalize treatment. The unique feature of MRgRT is that it permits daily MRI acquisition, creating possibilities for the development of models that predict tumor progression. This opens the door to considering biological/physiological changes within the tumor to predict outcomes.

In this work, a delta-radiomics model for PFS event prediction in patients receiving MRgRT was developed and evaluated. This includes a binary classification of the presence of a PFS event or not, which is a combined end point that includes local progression, distant progression, or death.

In this study, two delta-radiomics models (F1||F5 and F5/F1) were evaluated with and without the auxiliary lung and pancreatic data. We compared the performance of the decision tree classifier on the F1|F5 model and the F5/F1 model (Table 3) for optimal model selection. Considering the model training, for the F1||F5 model, with the lung data set, the AUC-ROC increased from 0.53 ± 0.12 to 0.61 ± 0.11, whereas the inclusion of the pancreatic data increased to 0.60 ± 0.14. For the F5/F1 model, the inclusion of the lung auxiliary data increased our AUC-ROC from 0.52 ± 0.13 to 0.65 ± 0.11, whereas it was 0.63 ± 0.13 when the pancreatic data were included.

During external testing (Table 4), for the F5/F1 model, we reported an AUC-ROC of 0.60 with the lung auxiliary data and 0.43 with the pancreatic data. Also, for the F5||F1 model, we reported an AUC-ROC of 0.70 with the lung auxiliary and 0.60 with the pancreatic data. Therefore, our optimal model is the F1||F5 model, which implies that the random selection of features across fractions (F1||F5) is more robust and performs better in the prediction of the external data set with the inclusion of the auxiliary data sets than predefined relationships between features from each fraction (F5/F1) in this case. The delta-radiomics model (F5/F1) accounts for the fractional physiological/biological changes within the tumor and can open doors for prospective clinical interventions during clinical trials, whereas the F1||F5 model accounts for random selection of features in the first and last fractions whose combination provides information that quantifies the changes within the tumor. The F5||F1 delta-radiomics model with the auxiliary lung data performed best on the external data set with an AUC-ROC of 0.70. This may be due to the fact that the majority of our external adrenal patients have primary lung cancer. The inclusion of pancreatic cancer with an AUC-ROC of 0.60 provided just enough information to classify the patients fairly. The lung data set provided significant information that improves the capacity of the model to classify the patients. The selected features of the F1||F5 model with the lung auxiliary data, its correlation map, and feature importance ranking can be found in the Data Supplement (Table S1 and Figs S1 and S2, respectively). Also, decision tree classifiers are known for interpretability and explainability of how the selected features are used for decision making. Our decision tree model gives not only the best predictive features but also the conditions/rules used for the classification of patients into those with progression events or not (Fig 2).

Previous studies have tried to use delta-radiomics in the low-field MRL. Cusumano et al²⁰ reported a multicenter delta-radiomics model for rectal cancer response prediction using low-field MRgRT. They identified two delta-radiomics features, the variation of least length and the gray-level nonuniformity, as predictors of clinical complete response and pathological clinical response. Although the primary end point was likely more reflective of response to treatment than the overall PFS used in this study, their model was trained on a smaller sample size compared with ours, and no signal normalization was performed on the MRIs across the different institutions. We accounted for the differences in the internal and external MRIs by performing fixed-bin number (FBN) intensity normalization on the gross tumor volume (GTV) using 64 bins. FBN intrinsically introduces a normalization effect by mapping the arbitrary intensity units of the raw MRI data into the same range for all patients and GTV while preserving the image contrast.²¹ As the values of some radiomics features depend on the number of gray levels found within a given GTV, using FBN gave room for directly comparing feature values across GTVs at different time points and samples.

Our group, Tomaszewski et al,²² also applied delta-radiomics to predict treatment in pancreatic cancer. Similar to this study, the ratio of the features between the first and last fractions (F5/F1) was used to predict PFS. Very few studies involve the application of radiomics to predict clinical outcomes using MRIs of adrenal cancers.^23,24 Stanzione et al²⁴ reported classifying adrenal lesions as benign or malignant using MRIs. The limitation is that it is a single-center analysis and does not account for changes in the radiomics features between fractions. However, similar to our study, GLSZM small area low gray level emphasis was one of the predictive radiomics features for adrenal lesion classification.

Some major obstacles to delta-radiomics biomarker development are the differences in imager settings and signal drift over time. As these low-field MRLs need very specific settings to function, differences in imaging settings are minimal across institutions. To assess radiomics signal drift, our group has shown the robustness of images from 0.35 T over various radiomics features.¹⁹ GLCM inverse difference, one of the features predicted in this study, is part of the reported robust and predictive features. With specific settings needed across institutions and robustness of signals assessed over time, these signal changes seen in tumors with daily MRI help us move toward imaging biomarker development. Our study is unique in this research space with the application of ML classifiers and the inclusion of auxiliary data sets to improve the efficiency of our model. We overcame overfitting due to a small data set by including an auxiliary data set during the model training, thereby producing a multidisease site model. The internal adrenal metastasis data set has 13 primary sites, whereas the external data set has three primary sites, with the lung being the predominant primary tumor. The inclusion of auxiliary data in the training data is a form of transfer learning. The goal is to enrich the adrenal training data with other related cancer site data. We learned similar imaging characteristics, such as texture and shapes, by taking advantage of the biological relationship between these sites. We chose lung cancer as our auxiliary data because over 80% of our external data set has primary lung cancer. We tested this hypothesis of enriching our training data with auxiliary data based on the biological relationship by including the pancreatic data. We provided examples of patients in the external data who were correctly/incorrectly predicted with and without the inclusion of the auxiliary data with respect to the site of their primary tumor (Data Supplement, Section S5 and Figs S6-S10).

To the best of our knowledge, we are the first group to identify delta-radiomics features that predict tumor progression during MRgRT that are reproducible across multiple institutions. With this, we ascertain the generalizability of our model based on its performance on the external data set. This is a critical hurdle that needs to be overcome for imaging biomarker development.

The limitation of this study is the sample size of our training data set, and the prediction of PFS event as a combined end point of local progression, distant progression, and death might be an imperfect measure to differentiate between patients with local or distant progression. The severity of class imbalance in the local failure result is the reason for the combined end point. We have eight of 42 internal adrenal patients and two of 18 external adrenal patients with local failure (Table 1). We overcame the small sample size using a novel approach of auxiliary data sets to train our model better. Our future work will be to increase our sample size across institutions, train our model with more auxiliary data sets depending on the prevalent primary tumor site, and consider using our model to predict the end point with the best prediction ability—local versus distant tumor progression. We would like to evaluate our model on a larger multicenter data set/multimanufacturer and multidisease site data set to continue to improve the generalizability of the model. The strength of our approach remains the external validation of an outside institutional data set and the use of auxiliary lung data, which has a biological relationship with the adrenal and the auxiliary pancreatic data, which does not have such a relationship. We further want to strengthen this work through prospective trials that use our model to predict tumor progression. If validated prospectively, this further opens the door for personalized biologically based radiotherapy treatment.

Given the reproducibility of the signal (even at low field), delta-radiomics features may provide important biological information about the tumor during treatment. There has been significant interest in looking at magnetic resonance-based biomarkers to predict tumor outcomes.²⁵ However, most efforts have focused on volumetric tumor changes and/or changes in signals within various sequences. These efforts have also been halted by the lack of frequency of MRI obtained. This is less of an obstacle with MRgRT as MRIs are obtained daily as a part of treatment. Although these signals may be predictive of PFS, the exact biological mechanism remains elusive at this time. We hypothesize this may be related to specific immune filtration, tumor necrosis/heterogeneity, and/or other biological processes.^26,27 We are currently working on clinical studies trying to correlate imaging-related changes to pathology to ascertain better how imaging/radiomics tumor habitat changes are related to biological processes (ClinicalTrials.gov identifier: NCT05301283).

In conclusion, developing a delta-radiomics model trained on SMART imaging data is feasible and was able to predict clinical outcome events related to disease progression. The F1||F5 model and the decision tree classifier are identified as the optimal delta-radiomics and ML classifier model, respectively, and achieved an AUC-ROC score of 0.70 on the external testing data, thus supporting the model’s generalizability. These results warrant further investigation using larger data sets. This opens a door for a feasibility study in biological dose painting of tumors during radiotherapy treatment.

CONTEXT.

Key Objective

To evaluate whether auxiliary data sets can improve the predictive performance of the machine learning model in using delta-radiomics features to predict disease progression in patients with adrenal metastasis.

Knowledge Generated

This study explores the use of delta-radiomics to predict progression-free survival events in patients receiving magnetic resonance-guided radiotherapy. By incorporating auxiliary data sets from patients with lung and pancreatic cancer, the predictive performance of the decision tree classifier was improved, especially with lung auxiliary data, since over 80% of the external adrenal patients have primary lung cancer.

Relevance

Temporally changing radiomics features may provide important early signals of both treatment response and risk of recurrence. This study is small but suggests that the technique has merit and should be evaluated on larger, multisite data sets.

DATA SHARING STATEMENT

A data sharing statement provided by the authors is available with this article at DOI https://doi.org/10.1200/CCI.24.00002 . The data may need additional IRB approval but will be available on request. The code can be found here—https://github.com/jfajemisin/Delta-Radiomics-Project.