A New 2.5D Representation for Lymph Node Detection using Random Sets of\n Deep Convolutional Neural Network Observations

Abstract

Automated Lymph Node (LN) detection is an important clinical diagnostic task\nbut very challenging due to the low contrast of surrounding structures in\nComputed Tomography (CT) and to their varying sizes, poses, shapes and sparsely\ndistributed locations. State-of-the-art studies show the performance range of\n52.9% sensitivity at 3.1 false-positives per volume (FP/vol.), or 60.9% at 6.1\nFP/vol. for mediastinal LN, by one-shot boosting on 3D HAAR features. In this\npaper, we first operate a preliminary candidate generation stage, towards 100%\nsensitivity at the cost of high FP levels (40 per patient), to harvest volumes\nof interest (VOI). Our 2.5D approach consequently decomposes any 3D VOI by\nresampling 2D reformatted orthogonal views N times, via scale, random\ntranslations, and rotations with respect to the VOI centroid coordinates. These\nrandom views are then used to train a deep Convolutional Neural Network (CNN)\nclassifier. In testing, the CNN is employed to assign LN probabilities for all\nN random views that can be simply averaged (as a set) to compute the final\nclassification probability per VOI. We validate the approach on two datasets:\n90 CT volumes with 388 mediastinal LNs and 86 patients with 595 abdominal LNs.\nWe achieve sensitivities of 70%/83% at 3 FP/vol. and 84%/90% at 6 FP/vol. in\nmediastinum and abdomen respectively, which drastically improves over the\nprevious state-of-the-art work.\n

Ähnliche Arbeiten