Ab­stract

Cap­sule en­doscopy is a method to cap­ture im­ages of the gas­troin­testi­nal tract and screen for dis­eases which might re­main hid­den if in­ves­ti­gated with stan­dard en­do­scopes. Due to the lim­ited size of a video cap­sule, em­bed­ding AI mod­els di­rectly into the cap­sule de­mands care­ful con­sid­er­a­tion of the model size and thus com­pli­cates anom­aly de­tec­tion in this field. Fur­ther­more, the scarcity of avail­able data in this do­main poses an on­go­ing chal­lenge to achiev­ing ef­fec­tive anom­aly de­tec­tion. Thus, this work in­tro­duces an en­sem­ble strat­egy to ad­dress this chal­lenge in anom­aly de­tec­tion tasks in video cap­sule en­do­scopies, re­quir­ing only a small num­ber of in­di­vid­ual neural net­works dur­ing both the train­ing and in­fer­ence phases. En­sem­ble learn­ing com­bines the pre­dic­tions of mul­ti­ple in­de­pen­dently trained neural net­works. This has shown to be highly ef­fec­tive in en­hanc­ing both the ac­cu­racy and ro­bust­ness of ma­chine learn­ing mod­els. How­ever, this comes at the cost of higher mem­ory usage and in­creased com­pu­ta­tional ef­fort, which quickly be­comes pro­hib­i­tive in many real-world ap­pli­ca­tions. In­stead of ap­ply­ing the same train­ing al­go­rithm to each in­di­vid­ual net­work, we pro­pose using var­i­ous loss func­tions, drawn from the anom­aly de­tec­tion field, to train each net­work. The meth­ods are val­i­dated on the two largest pub­licly avail­able datasets for video cap­sule en­doscopy im­ages, the Galar and the Kvasir-Cap­sule dataset. We achieve an AUC score of 76.86% on the Kvasir-Cap­sule and an AUC score of 76.98% on the Galar dataset. Our ap­proach out­per­forms cur­rent base­lines with sig­nif­i­cantly fewer pa­ra­me­ters across all mod­els, which is a cru­cial step to­wards in­cor­po­rat­ing ar­ti­fi­cial in­tel­li­gence into cap­sule en­do­scopies.