Tandem: segmentation of small medical objects exploiting classification

Deep Learning Medical Image Segmentation is a popular computer vision task, its goal is to provide a precise and accurate representation of target objects with the purpose of disease diagnosis or treatment planning. In this thesis we apply Deep Learning Image Segmentation methods to detect cysts on physiological images of kidneys tissues affected by ADPKD. The collected dataset is characterized by images depicting several sparse and tiny cysts with different sizes and shapes in order to improve segmentation results already computed in previous work over it. Moreover, given images and cysts characteristics we will focus our attention over deep learning methods developed to well perform also with images depicting really small objects. Different solutions will be explored and finally a proposed method consisting of a segmentation model and a classifier trained together, called Tandem method, will be presented and tested. Classification head role inside Tandem solution can be described as a post processing segmentation refinement specifically suitable for small and sparse target objects. Classifier head is aware of segmentation model predictions and consequently refine them by erasing the classified-as-wrong segmented object, in order to adjust and improve segmentation output. Computed tests over our dataset show that this method achieve great results, outperforming standard segmentation models such as U-Net, for example.