Radiomic-based machine learning algorithm to reliably differentiate benign renal masses from renal cell carcinoma

A non-transitory computer-readable medium comprising computer readable instructions, which when executed by a processor, cause the processor to perform operations for conducting a non-invasive, non-surgical, digital biopsy method for accurately predicting benign kidney lesions from cancers in a patient, the operations comprising:

• receiving, at the processor, a plurality of contrast-enhanced computed tomography image sets of a renal mass acquired in multiple CT phases, including at least a nephrographic phase image set and at least one additional phase image set;
• generating a three-dimensional region of interest corresponding to the renal mass by segmenting the renal mass in the nephrographic phase image set;
• co-registering the at least one additional phase image set to the nephrographic phase image set using a normalized mutual-information cost function to produce co-registered phase image sets;
• extracting, from voxel data within the region of interest in the co-registered phase image sets, a radiomics feature vector that includes shape features computed on an entire tumor volume and texture features derived from at least one of:
• gray-level histogram analysis, gray-level co-occurrence matrix analysis, gray-level difference/dependence matrix analysis, or frequency analysis based on a fast Fourier transform;
• applying to the radiomics feature vector a trained machine-learning classifier comprising a random forest classifier that (i) ranks variables of importance using an out-of-bag Gini index and (ii) was validated using tenfold cross-validation;
• and
• outputting, by the processor, a probability that the renal mass is malignant and a classification of the renal mass as benign or malignant based on the probability;
• wherein the trained machine-learning classifier is trained using a reduced radiomics panel that comprises radiomics features predetermined to be robust across multiple CT scanners and imaging protocols based on an intraclass correlation coefficient criterion.

A system for non-invasive, non-surgical, digital biopsy for accurately predicting benign kidney lesions from cancers in a patient, the system comprising:

• a diagnostic scan source comprising a computer-tomographic image scanner configured to generate a plurality of computer-tomographic images corresponding to at least one kidney of a patient;
• a server having a memory storing a plurality of patient clinical factors;
• and
• a processor connected to the diagnostic scan source and the server and configured to:
• receive at the processor a plurality of contrast-enhanced computed tomography image sets of a renal mass acquired in multiple CT phases, including at least a nephrographic phase image set and at least one additional phase image set;
• generate a three-dimensional region of interest corresponding to the renal mass by segmenting the renal mass in the nephrographic phase image set;
• co-register the at least one additional phase image set to the nephrographic phase image set using a normalized mutual-information cost function to produce co-registered phase image sets;
• extract, from voxel data within the region of interest in the co-registered phase image sets, a radiomics feature vector that includes shape features computed on an entire tumor volume and texture features derived from at least one of:
• gray-level histogram analysis, gray-level co-occurrence matrix analysis, gray-level difference/dependence matrix analysis, or frequency analysis based on a fast Fourier transform;
• apply to the radiomics feature vector a trained machine-learning classifier comprising a random forest classifier that (i) ranks variables of importance using an out-of-bag Gini index and (ii) was validated using tenfold cross-validation;
• and
• output, by the processor, a probability that the renal mass is malignant and a classification of the renal mass as benign or malignant based on the probability;
• wherein the trained machine-learning classifier is trained using a reduced radiomics panel that comprises radiomics features predetermined to be robust across multiple CT scanners and imaging protocols based on an intraclass correlation coefficient criterion.