Ab­stract

This paper pre­sents a method to ef­fi­ciently clas­sify the gas­troen­tero­logic sec­tion of im­ages de­rived from Video Cap­sule En­doscopy (VCE) stud­ies by ex­plor­ing the com­bi­na­tion of a Con­vo­lu­tional Neural Net­work (CNN) for clas­si­fi­ca­tion with the time-se­ries analy­sis prop­er­ties of a Hid­den Markov Model (HMM). It is demon­strated that suc­ces­sive time-se­ries analy­sis iden­ti­fies and cor­rects er­rors in the CNN out­put. Our ap­proach achieves an ac­cu­racy of 98.04% on the Rhode Is­land (RI) Gas­troen­terol­ogy dataset. This al­lows for pre­cise lo­cal­iza­tion within the gas­troin­testi­nal (GI) tract while re­quir­ing only ap­prox­i­mately 1M pa­ra­me­ters and thus, pro­vides a method suit­able for low power de­vices.