Fig 1.
The process of cropping a 256 × 256 thorax bone SPECT image from the original 256 × 1024 whole-body bone SPECT image.
Fig 2.
Curve fitting based technique for identifying thorax area.
a) The original curve and its fitted one; and b) The curves of the first and second derivatives.
Fig 3.
An example of preprocessing thorax bone SPECT image.
a) The original thorax bone SPECT image; b) The horizontally mirrored image; c) The horizontally translated image by + 6 pixels; and d) The rotated image by +5°.
Table 1.
An overview of the used data of SPECT images.
Fig 4.
Labelling SPECT image using the LableMe based annotation system.
Fig 5.
The architecture of the U-Net based segmentation network with shortcut module.
Fig 6.
The structure of a residual module.
Fig 7.
The architecture of the Mask-RCNN based segmentation network with spatial attention module.
Fig 8.
The structure of spatial attention module.
Fig 9.
An example of K-means clustering based hotspot segmentation with thorax bone SPECT image with K = 5.
Fig 10.
PA and loss curves of two segmentation models.
a) U-Net; and b) Mask R-CNN.
Table 2.
Experimental results on evaluation metrics for 2 280 samples of thorax bone SPECT imaging.
Fig 11.
The IoU values obtained by K-means based segmentation model for different K.
Fig 12.
A comparison of K-means based (green), U-Net based (blue) and the manually labelled (purple) segmentation results.
Fig 13.
A visualization of hotspots segmented by U-Net-Res model for two thorax bone SPECT images (the model segmented results marked with green and the manually labeled ones marked with red).
a) The best case; and b) The worst case.