Cataract　is　one　of　the　most　common　eye　diseases,　which　occupies　4.2%　of　the　population　all　over　the　world.　Automatic　cataract　detection　not　only　can　help　people　prevent　visual　impairment　and　decrease　the　possibility　of　blindness　but　also　can　save　the　medical　resources.　Previous　researchers　have　achieved　automatic　medical　images　detection　using　the　Convolution　Neural　Network　(CNN),　which　may　be　non-transparent,　unexplained　and　doubtful.　In　this　paper,　we　propose　a　novel　idea　of　interpretable　learning　for　explaining　the　result　of　cataract　detection　generated　by　CNNs,　which　is　a　result-oriented　explanation.　The　AlexNet-CAM　and　GoogLeNet-CAM　are　reestablished　on　basis　of　AlexNet　and　GoogLeNet　by　replacing　two　fully-connected　layers　with　global　average　pooling　layer.　Four　models　are　used　to　test　whether　class　activation　mapping　(CAM)　make　the　accuracy　dropped.　Then,　we　use　gradient-class　activation　mapping　(Grad-CAM)　combined　with　existed　fine-grained　visualization　to　generate　heat-maps　that　show　the　important　pathological　features　clearly.　As　a　result,　the　accuracy　of　AlexNet　(GoogLeNet)　is　94.48%　(94.89%),　and　that　of　AlexNet-CAM　(GoogLeNet-CAM)　is　93.28%　(94.93%).　Heat-maps　corresponding　with　non-cataract　fundus　images　highlighted　the　lens　and　parts　of　big　vessels　and　small　vessels;　and　the　clarity　of　three　kinds　of　heat-maps　corresponding　with　cataract　images　declined　in　turn,　which　are　mild,　medium　and　severe.　The　results　prove　our　approaches　can　keep　the　accuracy　stable　and　increase　the　interpretability　for　cataract　detection,　which　also　can　be　generalized　to　any　fundus　image　diagnosis　in　the　medical　field.　©　2019,　Springer　Nature　Singapore　Pte　Ltd.