Computer Vision News - October 2021

39 Teodora Popordanoska Best of MICCAI 2021 The second part of the title refers to volume estimation . In medical image analysis, the segmentation of an image is usually performed with a neural network and is mostly used to calculate certain biomarkers. In the medical domain, the volume of a tumor, organ, or lesion are important biomarkers. From a segmentation, one can obtain the volume by summing up the probability scores for each voxel . An important quantity in this case is the bias of volume estimate. Ideally, to have the true volume, there would be zero bias. “ The main theoretical result from this paper is that the absolute value of the bias is upper bounded by the calibration error, ” Teodora clarifies. “ If we’re optimizing calibration, we’re also simultaneously reducing the bias of the volume estimate. If the calibration error is zero, then it means that T e sec art f t e title refers t v l e esti ati . I e ical i age a alysis, t e seg e tati f a i age is s ally erf r e it a e ral et rk a is stly se t calc late certai i arkers. I t e e ical ai , t e v l e f a t r, rga , r lesi are i rta t i arkers. Fr a seg e tati , e ca tai t e v l e y s i g t e r a ility sc res f r eac v xel . i rta t a tity i t is case is t e ias f v l e esti ate. I eally, t ave t e tr e v l e, t ere l e zer ias. “ T e i t e retic l res lt fr t is er is t t t e s l te v l e f t e i s is er e y t e c li r ti err r, ” Te ra clarifies. “ If e’re ti izi c li r ti , e’re ls si lt e sly re ci t e i s f the volume estimate. If the calibration error is zero, then it means that