Unraveling response to temozolomide in preclinical GL261 glioblastoma with MRI/MRSI using radiomics and signal source extraction

Abstract

Glioblastoma is the most frequent aggressive primary brain tumor amongst human adults. Its standard treatment involves chemotherapy, for which the drug temozolomide is a common choice. These are heterogeneous and variable tumors which might benefit from personalized, data-based therapy strategies, and for which there is room for improvement in therapy response follow-up, investigated with preclinical models. This study addresses a preclinical question that involves distinguishing between treated and control (untreated) mice bearing glioblastoma, using machine learning techniques, from magnetic resonance-based data in two modalities: MRI and MRSI. It aims to go beyond the comparison of methods for such discrimination to provide an analytical pipeline that could be used in subsequent human studies. This analytical pipeline is meant to be a usable and interpretable tool for the radiology expert in the hope that such interpretation helps revealing new insights about the problem itself. For that, we propose coupling source extraction-based and radiomics-based data transformations with feature selection. Special attention is paid to the generation of radiologist-friendly visual nosological representations of the analyzed tumors.

Figure 1

Left, schematic representation of a mouse brain and slice positioning. Right, T2w MRI image for case C234, with the $32\times 32$ grid superimposed over it. Spectra colored in green are the ones belonging to the VoI (yellow square) of $10\times 10$ or $12\times 12$ enclosing the data used in the study. The rest of spectra are represented in red. Blue lines separate all grid pixels.

Figure 2

Example of a 16-levels Minkowski thresholds over a tumor mask (C526, treated case, day 18).

Figure 3

Example of how voting systems work on mouse C1320, control case. (a, left) Slice-voting system (SVS); (b, right) Voxel-voting system (VVS).

Figure 4

Example of how only spectra regions that are completely inside the tumor (blue outline) are selected from all the grid (red outline). Mouse C234.

Figure 5

Top: representation of performance (accuracy over the hold-out set) of Logistic Regression over the number of Radiomics features selected (n most relevant features according to t-test, with $n=1,\dots ,30$) for MRI data in (a) and (b). Slice classification method was used in (a) whyle SVS method was used in (b). Bottom: performance of SVM classifier with linear kernel (accuracy for the hold-out set) over the number of sources selected for the MRSI data in (c) and (d). Voxel classification method was used in (c), while VVS method was used in (d).

Figure 6

Examples of nosological visual representation of the classification results for MRSI data from their extracted sources. They are meant to provide radiologists with an intuitive interpretation tool. They consist of horizontal T2w MRI images of GL261 GB afflicted mice at different treatment/evolution day, superimposed with representative nosological maps of the classification reliability in different tumor regions for the SVM classifier with linear kernel and 10 features (sources) selected by t-test. The color-coding (see scale on the right) shows how reliable the model classification output can be considered and it represents a classifier output posterior probability. The lighter the color, the more reliable and vice versa. The red contour over some of the voxels represents those that were misclassified by that model. (a) C1465-day15, (b) C1109-day11, (c) C1412-day23, (d) C1474-day14, (e) C1320-day18, (f) C1026-day23. The color bars at the bottom represent the true class of the case, whereas the color bars at the top represent the percentage of voxels classified either as treated or as control for each case.

Figure 7

Examples of control (left, mouse C583) and treated, transiently responding to TMZ according to histopathological parameters (right, mouse C574) murine GL261 GB tumors. Note the appearance of hypointense zones (red circles) in T2w MRI from the treated mouse, noticeably different from the more homogeneous appearance observed in the control case.

Figure 8

Examples of two selected sources relevant for distinguishing spectra from control and treated, responding murine GL261 GB tumors. Some relevant metabolites have been indicated [Polyunsaturated fatty acids (PUFA), Lactate (Lac), Alanine (Ala) and the signals overlapped from Glutamate-glutamine, alanine and glycine (labeled as Glx)]. The green source shows some of the typical features seen in MRSI pattern of treated, responding tumors such as visible PUFA and relative increase in Lac, while the red source did not show (or only barely visible) PUFA, in addition to increased Glx signals and a more clear Ala presence in comparison with the green source.