Pathological lung segmentation (PLS) is an important, yet challenging,medical image application due to the wide variability of pathological lungappearance and shape. Because PLS is often a pre-requisite for other imaginganalytics, methodological simplicity and generality are key factors inusability. Along those lines, we present a bottom-up deep-learning basedapproach that is expressive enough to handle variations in appearance, whileremaining unaffected by any variations in shape. We incorporate the deeplysupervised learning framework, but enhance it with a simple, yet effective,progressive multi-path scheme, which more reliably merges outputs fromdifferent network stages. The result is a deep model able to produce finerdetailed masks, which we call progressive holistically-nested networks(P-HNNs). Using extensive cross-validation, our method is tested onmulti-institutional datasets comprising 929 CT scans (848 publicly available),of pathological lungs, reporting mean dice scores of 0.985 and demonstratingsignificant qualitative and quantitative improvements over state-of-the artapproaches.
展开▼
