Improving RetinaNet for CT lesion detection with dense masks from weak RECIST labels

Accurate, automated lesion detection in Computed Tomog-raphy (CT) is an important yet challenging task due to the large variationof lesion types, sizes, locations and appearances. Recent work on CT le-sion detection employs two-stage region proposal based methods trainedwith centroid or bounding-box annotations. We propose a highly accu-rate and efficient one-stage lesion detector, by re-designing a RetinaNetto meet the particular challenges in medical imaging. Specifically, we op-timize the anchor configurations using a differential evolution search al-gorithm. For training, we leverage the response evaluation criteria in solidtumors (RECIST) annotation which are measured in clinical routine. Weincorporate dense masks from weak RECIST labels, obtained automat-ically using GrabCut, into the training objective, which in combinationwith other advancements yields new state-of-the-art performance. Weevaluate our method on the public DeepLesion benchmark, consisting of32,735 lesions across the body. Our one-stage detector achieves a sensitiv-ity of 90.77% at 4 false positives per image, significantly outperformingthe best reported methods by over 5%