“…At another end, new methods will emerge to provide better representations for the encoded inputs via concepts like networks deconvolution, inversion, and dissection, among others (see [ 95 ]). Fifth and last, in order to face the challenge of intratumor heterogeneity, the quantification of tumor abundance at the voxel level is becoming an important direction in response assessment and recurrence risk studies [ 4 , 95 , 96 , 97 ]. This might help the identification of subregions, for instance, those metabolically active and defined as high risk [ 92 ], and may also inspire strategy to mitigate the effects of unbalanced data (for instance, when an outcome is over-represented) and thus decisional bias [ 12 ].…”