Background　Detecting　lesions　in　esophageal　cancer　(EC)　is　important　in　guiding　subsequent　treatment.　Deep　learning　methods　based　on　convolutional　neural　networks　(CNNs)　and　vision　transformer　(ViT)　have　made　remarkable　strides　in　the　field　of　medical　image　analysis　due　to　their　powerful　representational　capabilities.　However,　without　prior　knowledge,　both　traditional　CNNs　and　ViTs　are　susceptible　to　disregarding　critical　anatomical　information,　including　loops　and　voids.　The　current　methods,　combined　with　persistent　homology　(PH),　have　been　proposed　to　address　certain　limitations,　but　they　neglect　the　inconsistencies　among　features　caused　by　the　lack　of　interaction　between　features,　ultimately　leading　to　a　reduction　in　the　model＇s　generalization　ability.　Purpose　To　address　these　challenges,　we　propose　a　novel　framework,　combined　with　PH　and　feature　interaction,　for　identifying　EC　lesions　from　3D　CT　images.　The　goal　is　to　enhance　the　predictive　capability　of　existing　deep　learning　models　by　incorporating　both　topological　information　from　PH　and　effective　feature　interaction　mechanisms.　Methods　We　applied　cube-wise　classification　techniques　to　improve　the　detection　of　lesions　associated　with　EC.　The　proposed　framework　consists　of　two　fundamental　modules:　(1)　persistence　diagram　cross-attention　encoder　(PDCAE)　that　completely　encodes　the　persistence　diagram　(PD)　created　by　PH　through　cross-attention.　(2)　recalibration　guidance　module　(RGM)　connecting　the　PH　features　with　the　image　features　efficiently　to　remove　inconsistencies.　Results　The　experimental　results　show　that　the　proposed　modules　significantly　enhance　the　predictive　capability　of　standard　backbone　networks,　and　outperform　the　state-of-the-art　classification　network.　Conclusions　This　work　highlights　the　potential　of　combining　topological　data　analysis　with　deep　learning　for　medical　image　analysis　tasks.　More　potential　downstream　tasks　that　can　utilize　topological　relationships　remain　to　be　explored　in　the　future.