First let me mention a minor point concerning terminology. The type of channel you are suggesting is often called a Pauli channel; the term depolarizing channel usually refers to the case where $p_x = p_y = p_z$.
Anyway, it is not really correct to say that Pauli channels are the channel model considered for quantum error correction. Standard quantum error correcting codes can protect against arbitrary errors (represented by any quantum channel you might choose) so long as the errors do not affect too many qubits.
As an example, let us consider an arbitrary single-qubit error, represented by a channel $\Phi$ mapping one qubit to one qubit. Such a channel can be expressed in Kraus form as
$$
\Phi(\rho) = A_1 \rho A_1^{\dagger} + \cdots + A_m \rho A_m^{\dagger}
$$
for some choice of Kraus operators $A_1,\ldots,A_m$. (For a qubit channel we can always take $m = 4$ if we want.) You could, for instance, choose these operators so that $\Phi(\rho) = |0\rangle \langle 0|$ for every qubit state $\rho$, you could make the error unitary, or whatever else you choose. The choice can even be adversarial, selected after you know how the code works.
Each of the Kraus operators $A_k$ can be expressed as a linear combination of Pauli operators, because the Pauli operators form a basis for the space of 2 by 2 complex matrices:
$$
A_k = a_k I + b_k X + c_k Y + d_k Z.
$$
If you now expand out the Kraus representation of $\Phi$ above, you will obtain a messy expression where $\Phi(\rho)$ looks like a linear combination of operators of the form $P_i \rho P_j$ where $i,j\in\{1,2,3,4\}$ and $P_1 = I$, $P_2 = X$, $P_3 = Y$, and $P_4 = Z$.
Now imagine that you have a quantum error correcting code that protects against an $X$, $Y$, or $Z$ error on one qubit. The usual way this works is that some extra qubits in the 0 state are tacked on to the encoded data and a unitary operation is performed that reversibly computes into these extra qubits a syndrome describing which error occurred, if any, and which qubit was affected.
Supposing that the arbitrary error $\Phi$ happened on the first qubit for simplicity, after the syndrome computation you will end up with a state that looks like a linear combination of terms like this:
$$
P_i |\psi\rangle \langle \psi| P_j \otimes |P_i\: \text{syndrome}\rangle\langle P_j\:\text{syndrome}|.
$$
The assumption here is that $|\psi\rangle$ represents the encoded data without any noise, $P_i$ and $P_j$ act on the first qubit, and that "$P_i$ syndrome" and "$P_j$ syndrome" refer to the standard basis states that indicate that these errors have occurred on the first qubit. (The situation is similar for the error affecting any other qubit; I'm just trying to keep the notation simple by assuming the error happened to the first qubit.)
Now the key is that you measure the syndrome to see what error occurred, and all of the cross terms disappear because of the measurement. You are left with a probabilistic mixture of states that look like
$$
P_i |\psi\rangle \langle \psi| P_i \otimes |P_i\: \text{syndrome}\rangle\langle P_i\:\text{syndrome}|.
$$
The error is corrected and the original state is recovered. In effect, by measuring the syndrome, you "project" or "collapse" the error to something that looks like a Pauli channel.
This is all described (somewhat briefly) in Section 10.2 of Nielsen and Chuang.