Isoform inference is not straight forward, but is implemented in the cufflinks package. It's rather hard to distinguish to which transcript reads belong with short reads. More straightforward is to count on the exon level, less ambiguous. This you can do using DEXSeq
Depends on how you obtained your "raw isoform counts". If you have a very simple experiment, and I mean 2 groups to compare against, then Kallisto -> Sleuth is the most straightforward was to achieve what you want. Sleuth is however an absolute pain if you have a more complex design than two groups, there seems to be no support for contrast matrices like you'd see in Limma, or ways of interrogating your model fit with flexibility, like DESeq2.
This questions has been previously discussed at length here: Is There Any Other Method For Isoform-Level Differential Expression Analysis Except Cufflinks/Cuffdiff?
If your quantification technique worked well and you got the counts.
You can use the raw counts as input. There are many tools available for DE, example edgeR, DESeq2, DEXseq etc,
Or else there are specialized method to detect differential Isoforms. Like mentioned in this following post.
I agree fanli.gcb, there are no perfect method to perform isoform quantification and differential expression. But there are studies uses available methods to perform the task. There is also a recent study where they used similar approach as I described in previous post using MISO,
And here is the recent study,