|
| Algorithms | Advantages | Disadvantages |
|
| Watershed [225] | Being able to divide an image into its components | Takes too long to run in order to meet the deadline, sensitivity to false edges and over-segmentation |
| U-Net [226] | Images can be segmented quickly and accurately | Redundancy occurs due to patch overlap, also relatively slow |
| MV-CNN [203] | No user-interactive parameters or assumptions about the shape of nodules are needed | The loss of gradients may have an effect |
| CF-CNN [206] | Gathered sensitive information about nodules from CT imaging data | Less adaptable for small nodules and cavitary nodules |
| FCM [188] | Ignored noise sensitivity limitation, successfully overcoming the PCM’s clustering problem | Row sum constraints must be equal to one in order to work |
| Hessian-based approaches [209] | High robustness against noise and sensitivity to small objects | Performance decreases for large nodule |
| SegNet + shape driven level set [213] | Correct seed point initialization with no manual intervention in the level set | Segments the lung nodule partly occluded, also takes a longer time |
| Faster R-CNN [214] | The efficiency of detection is high | It could take a long time to reach convergence |
| Mask R-CNN [218] | Easy to train, generalizable to other tasks, effective, and only adds a minor overhead | Low-resolution motion blur detection typically fails to pick up on objects |
| RASM [219] | Well suited to large shape models and parallel implementation allowing for short computation times | Cannot segment areas with sharp angles and is not built to handle juxta-pleural nodules |
| Region growing [227] | The concept is simple, multiple criteria can be selected simultaneously, and it performs well in terms of noise | Computing is time-consuming. Noise or variation may result in holes or over-segmentation, making it difficult to distinguish the shading of real images. |
|
