Reframing Universal Lesion Segmentation as a Large-Organ Lesion Segmentation Task
Keywords: Universal lesion detection, Lesion segmentation, Crowd-sourcing image
Abstract: Universal lesion segmentation is challenging due to (1) the need to segment lesions across the entire body, often when they occupy only a small portion of the image, and (2) the crowdsourced nature of training images, leading to inconsistent annotation quality. Many existing methods adopt a divide-and-conquer strategy or integrate detection with segmentation to enhance training effectiveness. In contrast, we simplify the task by treating all lesions as a single type and directly training a universal lesion segmentation model using large image spacing and input volumes. Our approach is inspired by single-organ tumor segmentation, where including a large portion of the organ improves performance. Extending this concept, we consider the entire human body as the "organ" for universal lesion segmentation. However, applying conventional settings for single-organ segmentation to the whole body is computationally expensive and requires substantial GPU memory. To address this, we employ large volume spacing during training, effectively balancing model complexity, training cost, and performance. Our method achieved a Dice score of 0.66 and an NSD of 0.59 on the online validation set and a Dice score of 0.46 and an NSD of 0.38 on the test set, ranking first on both leaderboards. The inference time is approximately 80 seconds per case, with a GPU memory requirement of 4 GB.
Submission Number: 12
Loading