Fig 1.
Image processing workflow for correlating MRI with histopathology and predicting tumor necrosis using MRI.
Each dashed box specifies the investigated parameters in association with their respective steps in the workflow.
Table 1.
MR images of different modalities and their derived parametric maps.
Table 2.
Statistical parameters and Haralick texture features computed from MR images.
Table 3.
Demographic information for patients in the study (patients included in the imaging data analysis are highlighted).
Fig 2.
Co-registration of histologic and MR images of osteosarcoma in POI through control point mapping.
Representative images are shown for a good-responding patient (Nhistℓ ≥ 90%) and a poor-responding patient (Nhistℓ < 90%). (A) Post-contrast T1-weighted MR image, with the boundary of the tumor AOI drawn in red. (B) Histologic section image (reconstructed from individual whole slide images) prior to image registration, and (C) its corresponding histologic tumor viability map (wherein necrotic and viable tumor regions are indicated in red and blue, respectively) computed by a deep learning classification model. Each pair of defined control points is represented by a yellow dot (in MR image) and a blue dot (in histologic image) labeled with the same number. (D) Histologic image and (E) its associated tumor viability map after being co-registered to MR image. (F) Average intensity projection of MR image fused with histologic image. (G) Average intensity projection of MR image fused with histologic tumor viability map.
Fig 3.
Statistical significance of difference in MR image features between histologic necrosis and viable tumor.
W and G denote the window size and the number of gray levels, respectively. Features are identified by their numbering in Table 2. Light blue indicates P < 0.001 (highly significant), medium blue indicates 0.001 ≤ P < 0.05 (significant), and dark blue indicates P ≥ 0.05 (not significant).
Fig 4.
Effects of window size W and number of gray levels G on Haralick texture features.
Images show the differences in texture features, such as contrast, entropy, and homogeneity, computed for the tumor AOIhist in the post-contrast T1-weighted MR image as window size W varies (from 3 to 9 to 15 pixels) while keeping the number of gray levels constant at G = 100. No visible differences can be observed in the texture features when the number of gray levels G changes from 50 to 100 to 200 while the window size remains constant at W = 9 pixels.
Fig 5.
Optimal values of objective function for different combinations of weight optimization methods and MRI subsets.
Optimal values of objective functions favg and fmax represent the minimized mean absolute error and the minimized maximum absolute error, respectively, obtained in the weight optimization process.
Fig 6.
Absolute error in tumor necrosis estimated using various combinations of weight optimization methods and MRI subsets.
The leftmost box represents the absolute error in histologic tumor necrosis estimated from primary histologic section images using a previously established deep learning classification method.
Fig 7.
Viability mapping of tumor AOI computed using different combinations of weight optimization methods and MRI subsets.
Tumor viability maps are shown for a good-responding patient (Nhistℓ ≥ 90%) and a poor-responding patient (Nhistℓ < 90%). Necrotic and viable tumor regions are represented in red and blue, respectively.
Fig 8.
Measures of similarity of necrotic tumor region between tumor viability map computed by histopathology and that estimated from various MRI subsets.
(A) Dice coefficient. (B) Overlap coefficient.
Fig 9.
Viability maps for tumor VOI estimated using different combinations of weight optimization methods and MRI subsets.
Three-dimensional tumor viability maps are shown in the coronal orientation for a good responder (Nhistℓ ≥ 90%) and a poor responder (Nhistℓ < 90%).Necrotic and viable tumor regions are represented in red and blue, respectively.
Fig 10.
Comparison of volume- and single plane-based estimations of tumor necrosis.
(A) Actual difference in tumor necrosis estimation between that of VOI and AOI (i.e., NVOIℓ–NAOIℓ). (B) Actual difference between VOI-based and histopathologic tumor necrosis estimations (i.e., NVOIℓ–Nhistℓ). An asterisk below a box indicates that the associated actual difference is statistically significant (P < 0.05).