Automatic Detection and Segmentation of Lung Lesions using Deep Residual CNNs

Abstract

Early detection of lung cancer has shown to significantly improve patient survival. Apart from lesion detection, tumour segmentation is critical for developing radiomics signatures. In this work, we propose a novel hybrid approach for lung lesion detection and segmentation on CT scans, where the segmentation task is assisted by prior detection of regions containing lesions. For the detection task, we introduce a 2.5D residual deep CNN working in a sliding-window fashion, whereas segmentation is tackled by a modified residual U-Net with a weighted-dice plus cross-entropy loss. Experimental results on the LIDC-IDRI dataset and on the lung tumour task dataset within the Medical Segmentation Decathlon show competitive detection performance of the proposed approach (0.902 recall) and superior segmentation capabilities (0.709 dice score). These results confirm the high potential of simpler models, with lower hardware requirements, thus of more general applicability.