There's a case to be made for alternatives to MIT Scheme. A good example of this is the exercises where you might employ the OOP system that SICP describes like the constraint-solver or metacircular evaluator but realize that you could be much more productive if you had some kind of controlled method for generating similar piles of code from some base implementation, something which is found in the object systems like CLOS like GOOPS or ChrisH's SOS or even Emacs's EIEIO here [2].
There's no impediment to completing Chapters 1 & 2 in Clojure, but the first chapters are not what is called to mind by "Structure and Interpretation of Computer Programs" [1].
- Function redefinition is extremely useful required if you want to write your code in a linear fashion as intended. Clojure automatically sets up a major hindrance here.
- A few subchapters explore parallel computation, Clojure is likely to lead you astray here as you try to find directly corresponding functions and why exactly the authors are suggesting.
- Several chapters explore mutable lists in detail. Serious complaints have been made about the suitability of Racket here, where there exists no more of a barrier than that methods which operate on mutable lists are prefixed with a specially named qualifier. Clojure would be functionally impossible here.
- On the plus side, the "JIT" meta-evaluator will be a much more interesting challenge!
You'll notice that all of these issues can be worked around. However, SICP with exercises will take you about a year if you're a serious student.
[1] Huffman trees exercises excluded.
[2] Doing the whole thing in Emacs Lisp is potentially fantastic. I don't know if I can wholeheartedly recommend it, but I would be really pleased if someone wrote tests and built the metaevaluator, compiler and really went the whole nine yards here.