Automatic liver tumor segmentation in CT with fully convolutional neural networks and object-based postprocessing

Abstract

Automatic liver tumor segmentation would have a big impact on liver therapy planning procedures and follow-up assessment, thanks to standardization and incorporation of full volumetric information. In this work, we develop a fully automatic method for liver tumor segmentation in CT images based on a 2D fully convolutional neural network with an object-based postprocessing step. We describe our experiments on the LiTS challenge training data set and evaluate segmentation and detection performance. Our proposed design cascading two models working on voxel- and object-level allowed for a significant reduction of false positive findings by 85% when compared with the raw neural network output. In comparison with the human performance, our approach achieves a similar segmentation quality for detected tumors (mean Dice 0.69 vs. 0.72), but is inferior in the detection performance (recall 63% vs. 92%). Finally, we describe how we participated in the LiTS challenge and achieved state-of-the-art performance.

Figure 1

Overview of the neural network architecture. The numbers denote the feature map count.

Figure 2

Non-trivial output (dashed)/reference (solid) correspondences. (a) Reference tumor corresponds to two output tumors (b) Three reference tumors correspond to one output tumor (c) Output tumor corresponds only to the smaller reference tumor.

Algorithm 1

Establishing correspondences between output and reference tumors.

Figure 3

MTRA (dashed) vs. LiTS (solid) annotations. (a) Case with low dice/correspondence (b) Case where a LiTS reference tumor was missed (c) Case where MTRA found a lesion in a case with no tumors according to LiTS reference (d) Case where small additional tumors were found by the MTRA.

Figure 4

Box plots showing dice per case (a) an dice per correspondence (b) computed for expert and automatically generated segmentations on 30 test cases.

Figure 5

Neural network (black) compared with the LiTS (white) annotations. (a) Case with 0.85 dice/case (b,c) Cases with 19 and 16 FPs (d) Case where a small tumor was not detected (e,f) Case where tumor segmentation strongly differed on consecutive slices.