Fig 1.
The framework of the proposed small tissue lesions segmentation method.
Fig 2.
Samples of image after preprocessing.
(a) Original image of LUNA16. (b) Fig (a) after preprocessing. (c) Original image of LiTS17. (d) Fig (c) after preprocessing.
Table 1.
HU values of typical objects and organs.
Fig 3.
Distribution of tissue lesions in CT images.
(a) and (b) are pulmonary nodules, and (c) and (d) are liver tumours.
Fig 4.
(a) and (c) show the original images of pulmonary nodules, while (b) and (d) show postsampling examples corresponding to images (a) and (c) the sampling rate in these examples was set to 4.
Fig 5.
Specific process of multiple oversampling fusion augmentation method.
Fig 6.
Samples of multiple oversampling fusion augmentation method.
(a) original image of pulmonary nodule image, (b) is the result of (a) using fusion oversampling, (c) original image of liver tumor image, and (d) is the result of (c) using fusion oversampling. (a red circle represents the original lesion area, a yellow circle represents mirrored and rotated lesion area, a blue circle represents magnified and rotated lesion area, and a green circle represents magnified and mirrored lesion area).
Fig 7.
Pixel distribution of (a) original images and (b) after class weighting for the LUNA16 dataset.
Fig 8.
Pixel distribution of (a) original images and (b) after class weighting for the LiTS17 dataset.
Fig 9.
The detailed architecture of the Mask-RCNN.
Fig 10.
The detailed architecture of the U-Net.
Fig 11.
The detailed architecture of the SegNet.
Fig 12.
The detailed architecture of the DeepLabV3+.
Table 2.
Performance of pulmonary nodules segmentation on LUNA16 data sets.
Table 3.
Performance of liver tumor segmentation on LiTS17 data sets.
Table 4.
The running time of different methods in segmenting lesion.
Table 5.
Performance of lesion segmentation on LUNA16 and LiTS17 data sets.
Fig 13.
A visual comparison of the lesion segmentation results.
(a)t-aug. (b)sp-aug. (c)mix-aug. (d)mof-aug. The red and green contours denote the ground truth and the segmentation results, respectively.
Table 6.
Comparison of our work to pulmonary nodule segmentation state-of-the-art methods.
All results are multiplied by 1000 and the bold font highlights the best results.
Table 7.
Comparison of our work to liver tumour segmentation state-of-the-art methods.
All results are multiplied by 1000 and the bold font highlights the best results.
Fig 14.
Schematic diagram of anchor frame matching with pulmonary nodules.
(a) Anchor frames that matched images of pulmonary nodules in the baseline training set. (b) Anchor frames that matched images of pulmonary nodules in the training set after random oversampling.
Fig 15.
The images of lung nodules and liver tumors.
(a) images of lung nodules. (b) Liver tumor image.
Table 8.
Effectiveness of Mask RCNN to the lesion size (Dice:%).