Study on the Impact of Neural Network Architecture and Region of Interest Selection on the Result of Skin Layer Segmentation in High-Frequency Ultrasound Images

Computer-Aided Diagnosis Methods for High-Frequency Ultrasound Data Analysis: A Review

“…Based on Table 2 and Table 4 , Table 5 , Table 6 , Table 7 , Table 8 , Table 9 and Table 10 , which are presented in the next sections, as well as Figure 4 (summary of individual applications for CAD of skin), we concluded that the most widely explored area was skin layer segmentation [ 3 , 10 , 14 , 15 , 16 , 17 , 27 , 48 , 64 , 76 ]. However, the majority of the works limited the analysis to the epidermis region, which is often crucial for further automated processing steps but is not sufficient for a complete diagnosis.…”

“…The application of the newest U-shaped models, basic U-Net by Ronnenberger et al [ 66 ], DC-UNet [ 71 ], and CFPNet-M [ 72 ], to epidermis and SLEB segmentation can be found in [ 48 ]. The authors analyzed the influence of the size of the images used for network training, augmentation technique, optimization method, region of interest (ROI) selection, and binarization threshold on the final segmentation accuracy.…”

“…An interesting observation is connected with the augmentation step, according to which the limited rotation works better for this type of image (layered). The best results obtained by each of the models investigated in this work [ 48 ] are included in Table 5 .…”

“…The next work by Czajkowska et al [ 74 ] described epidermis segmentation as a part of the HFUS skin image classification framework. The analyzed image set was extended (compared to [ 3 , 48 ]) by 200 images of non-melanocytic skin tumors. However, because the SLEB layer was not present in patients with skin tumors, it was omitted in this analysis.…”

“…For this, they combined deep models with the fuzzy connectedness (FC) analysis developed by Udupa et al [ 107 ] for fine segmentation. Similar to [ 48 ], the authors considered the ROI selection step and the experimental results favored it in this case. The pre-processing step targeting the ROI selection utilized a DeepLab v3+ [ 73 ] network built on a pre-trained backbone for feature extraction.…”