Ensembling Voxel-Based and Box-Based Model Predictions for Robust Lesion Detection

This paper presents a novel generic method to improve lesion detection by ensembling semantic segmentation and object detection models. The proposed approach allows to benefit from both voxel-based and box-based predictions, thus improving the ability to accurately detect lesions. The method consists of 3 main steps: (i) semantic segmentation and object detection models are trained separately; (ii) voxel-based and box-based predictions are matched spatially; (iii) corresponding lesion presence probabilities are combined into summary detection maps. We illustrate and validate the robustness of the proposed approach on three different oncology applications: liver and pancreas neoplasm detection in single-phase CT, and significant prostate cancer detection in multi-modal MRI. Performance is evaluated on publicly-available databases, and compared to two state-of-the art baseline methods. The proposed ensembling approach improves the average precision metric in all considered applications, with a 8% gain for prostate cancer.