4.2 Inter-slice and sequence HAN
Our inter-slice and sequence HAN (hereafter sequence-HAN) was designed
for abnormality
detection using a collection of image slices from multiple MRI
sequences. Like the intra-
and inter-slice network, it built up a hierarchical data representation
for abnormality classification, but this time at the series level. This
was achieved by first building up individual sequence representations
using separate LSTM units, CNNs, and slice-level attention modules. A
sequence-level attention module then weighed the importance of each
sequence representation, and the resulting series-level representation
was again sent to a single layer classifier (Fig. 3).