A preoperative clue for pediatric Wilms tumor subtypes: D* separates stromal from epithelial, mixed and regressive slices

Preoperative MRI parameters in pediatric Wilms tumor can now be matched slice by slice to postoperative histopathology, and at the macroscopic slice level they separate stromal from epithelial, mixed and regressive components. The protocol does not biopsy before treatment, so histological subtyping and risk stratification happen only after surgery, leaving no histological label available preoperatively. Ninety-four slices from 30 lesions were analyzed: 27 stromal, 27 mixed, 30 regressive, 5 epithelial, 3 blastemal and 2 completely necrotic; in mixed model analysis median D* was significantly higher in stromal than in epithelial (p = 0.024), regressive (p = 0.031) and mixed (p = 0.043) slices.

Parameter estimation moves to implicit neural representations, which fold spatial correlation across the image into the fit instead of solving each voxel on its own. Voxel-wise least-squares or segmented fitting is unstable for f and D* at low signal-to-noise ratio, and D* is the parameter that carried the subtype separation here. On simulated data the largest gain appeared for D*, where the implicit neural representation reached a Pearson correlation of 0.81 and root-mean-square error of 13.74, against 0.62 and 24.38 for segmented fitting; fitting ran for 300 epochs on an NVIDIA Titan X with a mean-squared-error loss.

The findings rest on single-center retrospective data: 30 lesions drawn from 56 patients, 1.5-T scans with 10 b-values, and all lesions having received preoperative chemotherapy. Correlation was done at the macroscopic slice level, with cutting guides holding 5 mm slice thickness and reproducing the MRI-DWI orientation. The supported use is separating stromal from epithelial, mixed and regressive components, not replacing lesion-level diagnosis for a patient.

Only 5 epithelial and 3 blastemal slices were available, too few for definitive conclusions, and more slices are needed. The dataset is shaped by preoperative chemotherapy, with fewer blastemal and more mixed and regressive lesions, and chemotherapy-resistant epithelial and blastemal tissue may not represent chemotherapy-sensitive versions. Tumor deformation after resection may still introduce errors in the imaging-to-pathology comparison.