Purpose

Suspected RCCs are not routinely biopsied prior to surgical resection; therefore the presence of sarcomatoid de-differentiation (sRCC) is often unknown until histologic evaluation. sRCC is associated with a poor prognosis, and is therefore managed more aggressively than renal cell carcinoma without sarcomatoid components (nsRCC). There is a clinical need for a non-invasive method to detect sarcomatoid dedifferentiation pre-operatively, especially when considering management options like active surveillance [1]. We report a supervised machine learning approach to classify renal cell carcinoma (RCC) tumors based on their appearance on T2-weighted (T2W), non-contrast T1-weighted fat-saturated (T1W), contrast-enhanced arterial phase T1-weighted fat-saturated (T1W-CEart), and contrast-enhanced venous phase T1-weighted fat-saturated (T1W-CEven) MRI images. The model was trained on these multiparametric MRI (mpMRI) data from 9 nsRCC and 9 sRCC tumors, validated on a separate cohort of 3 nsRCC and 3 sRCC tumors, and tested on a held-out set of 5 nsRCC and 5 sRCC tumors.

Methods

In this IRB-approved study we identified 17 subjects with pathology-proven sRCC tumors for whom pre-operative T2W, T1W, T1W-CEart and T1W-CEven MRI scans were available. We also curated a set of 17 nsRCC tumors of comparable tumor volumes. Tumors were divided into balanced training, validation and test cohorts as in Table 1. The T1W, T1W-CEart and T1W-CEven images were resampled and spatially co-registered to the T2W image. The contralateral normal kidney cortex was semi-automatically segmented on all applicable slices of the T1W-CEart images to provide a large number of voxels for intensity calibration as shown in Figure 1 . The tumor was manually contoured on all applicable slices by a Radiologist as illustrated in Figure 1. Raw tumor voxel intensities on T2W images were calibrated by dividing them by the mean intensity of the contralateral normal kidney cortex on T2W of the same patient. Raw tumor voxel intensities on T1W, T1W-CEart and T1W-CEven images were calibrated by dividing them by the mean intensity of the contralateral normal kidney cortex on T1W of the same patient. Calibrated voxel intensities from tumors belonging to the training and validation cohorts were variance-normalized to the means and standard deviations of pooled training + validation tumor voxels on each respective scan type; the same means and standard deviations were also used to variance-normalize intensities of voxels in tumors from the test cohort. After this process each voxel had 4 “channels” of calibrated and variance-normalized mpMRI intensities associated with it [2-4]. Reasoning that each mpMRI “channel” may be adequately described by 3 intensity levels, for a total of 3⁴ = 81 levels, we employed a Kohonen Self-Organizing Map (SOM) [5] of order 9x9 to reduce this 4-dimensional intensity space to 81 levels.

Results

A 9x9 SOM was trained on the 18 training cohort tumors to produce the SOM depicted in Figure 2(a); the 81 “neurons” in the trained SOM correspond to distinct “phenotypes” in the 4-dimensional calibrated & variance-normalized intensity space. Figure 2(b) shows the Activation Map of all 9 nsRCC tumors in the training cohort, while Figure 2(c) shows the Activation Map of all 9 sRCC tumors in the training cohort; each cell in a given Activation Map represents the relative number of voxels belonging to a given “phenotype” in the corresponding SOM. Activation Maps were generated individually for each of the training, validation and test tumors, and these became the inputs on which a Learning Vector Quantizer [6] (LVQ) classifier was trained. The LVQ was trained on Activation Maps of the 18 training cohort tumors and optimized for step size and number of training epochs. The model with the best performance (learning rate = 0.001, 1000 training epochs, Figure 3) on Activation Maps of the 6 validation cohort tumors was advanced to the test phase. This model produced an accuracy of 70% on the 10 test cohort tumors (Figure 4).

Discussion & Conclusions

We have demonstrated a Self-Organizing Map based approach to analysis of standard multiparametric MRI images for the task of identifying sarcomatoid dedifferentiation in renal cell carcinoma. Sarcomatoid de-differentiation is noted in approximately 5-10% of all RCCs, and this relatively uncommon presentation, combined with our requirement for availability of pre-operative MRIs, restricted the current study to a relatively small sample size. Our ongoing work is focused on increasing the sample size as well as on increasing the number of mpMRI “channels” to increase the information available to the SOM-LVQ model. In particular, we are seeking to add Apparent Diffusion Coefficient maps that are known to be useful for differential diagnosis in this setting [1].

Acknowledgements

No acknowledgement found.