Dimension ReductionKnowledge ClusteringEmbeddingMulti-HeadAttentionFeed ForwardAdd & NormAdd & NormReport EmbeddingEmbeddingMasked Multi-HeadAttentionMulti-HeadAttentionFeed ForwardLinear & SoftMaxSentence BertCosine Similarity384CNNCNN7×7×204820484096731024TransformerEncoderTransformerDecoderSimilarityComparerAdd & NormAdd & NormAdd & NormGround Truth ReportHeartsizeisnormal,thelungsareclear.Nopleuraleffusionorpneumothorax…AslabelsOutputGround Truth ReportHeart size is normal, the lungs are clear. No pleural effusion or pneumothorax… Heart size is normal, the lungs are clear. No pleural effusion or pneumothorax… Heart size is normal, the lungs are clear. No pleural effusion or pneumothorax… Knowledge DistillerKnowledge Matched Visual ExtractorReport GeneratorTheheartsizeis…𝑬𝒊∈𝑹𝟑𝟖𝟒…𝒀𝒊∈𝑹𝟐𝒕𝟏,𝒕𝟐 ,...,𝒕𝟕𝟑𝑽𝒂𝒗𝒈𝑽𝒂𝒗𝒈′𝑽𝒍𝑽𝒂𝑽𝒍′𝑽𝒂′