In this paper, the multicasting of digitally encoded images on a heterogeneous network is considered. In order to obtain analytically tractable problems, the wavelet transform coefficients of a digital image are modeled as a set of parallel Gaussian sources. Also, a general network transport mechanism subject to packet losses is modeled as an erasure broadcast channel where users are affected by possibly very different erasure probabilities. In the proposed setting, the convex nature of the rate distortion function allows relevant optimization problems corresponding to various performance criteria to be solved. The solutions of these optimization problems serve as starting points for the design of source-channel codes based on embedded scalar quantization, on linear rateless encoders that map directly the (redundant) bits generated by the quantizer into channel symbols, and on progressive transmission of the encoded symbols organized into "layers", such that users with higher capacity achieve better end-to-end distortion. At the decoders, iterative belief propagation decoding, multi-stage sequential decoding of the layers and soft-bit reconstruction are used. Numerical experiment sshow that 1) the proposed model is sufficiently accurate to provide system design guidelines for the case of real-life images, and 2) the proposed coding scheme achieves rate-distortion performance very close to the theoretical optimum.