Fig 1.
Flowchart of patient selection.
Patients were selected from a prospectively maintained registry based on their imaging and treatment protocols. CT = computed tomography.
Fig 2.
Patients from our prospectively maintained registry were selected based on their imaging and treatment protocols. CT exams were collected. All liver metastases were manually segmented. For each metastasis, 3-D radiomics were extracted and indexed, allowing multiple aggregation strategies. The radiomics signatures were used to stratify the risk of tumor recurrence and patient survival. CT = computed tomography.
Fig 3.
Investigated aggregation strategies.
Considering multiple metastases, per lesion radiomic features are combined in six distinct schemes. Coefficients a,b,c and d indicate lesions volume ratios.
Table 1.
Clinical characteristics of 241 patients.
Fig 4.
Time to recurrence (TTR) prediction using survival models.
Concordance indexes obtained on holdout test dataset using survival models across aggregation strategies for TTR prediction with (A) random survival forests and (B) DeepSurv. Green dashed line indicates performance of CRS alone.
Fig 5.
Disease-specific survival (DSS) prediction using survival models.
Concordance indexes obtained on holdout test dataset using survival models across aggregation strategies for DSS prediction with (A) random survival forests and (B) DeepSurv. Green dashed line indicates performances of CRS alone.
Fig 6.
Time to recurrence (TTR) classification using radiomics signatures.
AUC obtained on holdout test dataset using radiomics signatures across aggregation strategies for TTR classification with (A) radiomics signature and (B) DeepSurv score. Green dashed line indicates performances of CRS alone.
Fig 7.
Disease-specific survival (DSS) classification using radiomics signatures.
AUC obtained on holdout test dataset using radiomics signatures across aggregation strategies for DSS classification with (A) radiomics signature and (B) DeepSurv score. Green dashed line indicates performances of CRS alone.
Fig 8.
Radiomics signature coefficients for prediction of recurrence.
(A) Box plot shows selected features coefficients sorted by descending order of selection ratio over splits. (B) Count plot of radiomics signature coefficients selections over the splits for the aggregation the radiomics signature coefficients with the aggregation ‘largest lesion only’.
Fig 9.
Liver metastases examples with their associated segmentations.
First row exhibits (A) initial CT image (left) and (B) a segmented metastasis (yellow area) of the patient with the shortest survival in our dataset. Second row exhibits (C) initial CT image (left) and (D) a segmented metastasis (yellow) of the patient with the longest survival in our dataset.