Leafcutter relies only on splice-junction-spanning reads, which allows it to be computationally very fast and scalable to 1000s of samples.
However, because Leafcutter does not include exon or intron coverage information, it is not suitable for detecting intron retention or other splicing events which involve coverage differences - UTR extensions etc.
The Leafcutter approach is to cluster overlapping splice junction reads, and to test for differential use of the reads that contribute to the cluster using a multinomial test. The simplest example would be an exon extension, which would have 2 different introns that overlap on only 1 end, while the other end would vary. The next example would be a cassette exon, with 3 introns: 1 intron spanning the adjacent exons which excludes the central cassette, and 2 introns that span from the adjacent exons into the cassette, which includes the central cassette.
Now Leafcutter doesn't know what a "cassette exon" is - it doesn't know that the introns must overlap only on certain ends, that the end of the one intron must be upstream of the start of another, etc. This is why I built a set of tests for cassette exons into the standalone version of Leafviz - https://github.com/jackhump/leafviz#classify-clusters-as-cassette-exons-and-determine-directionality-and-novelty
This script attempts to screen your leafcutter clusters for those that match our interpretation of a cassette exon and assigns directionality.
In my hands it has been useful to cut down a large set of differentially spliced clusters into a smaller and more actionable set of cassette exons, which I used in this paper: https://pubmed.ncbi.nlm.nih.gov/34274995/
I agree with the previous poster that are many different tools available to perform differential splicing analysis. I would say that Leafcutter is useful in the following scenarios:
1) You have 100s - 1000s of samples for testing - although RMATS and MAJIQ now have fast versions for this.
2) Your research question is based around the detection of novel splicing events - this question is also addressed by MAJIQ, but not by RMATS or SUPPA.
3) You want to take an intron-centric approach, as opposed to an isoform-centric approach which is typified by SUPPA.