Page 82 - AIH-2-2

Artificial Intelligence in Health Efficient knowledge distillation for breast US

terminology throughout this manuscript by referring to this et al. proposed different approaches for transfer learning.
71
dataset accordingly (refer to Section 3.1 for more details on In one of their experiments, first, they pre-trained various
the dataset). Using Dataset_A, Yap et al. proposed an end- networks on Achilles tendon US images, and then fine-
47
to-end approach for US lesion detection and recognition tuned on breast US images (i.e., Dataset_A) and reported
by utilizing a pre-trained segmentation network designed the best DSC of 83%. A unified-focal loss was introduced by
based on fully convolutional networks (FCN) and achieved Yeung et al. achieving a DSC of 82%. An adaptive receptive
48
72
Dice similarity coefficient (DSC) score of 55%. Abraham and field network proposed by Xu et al. reported a DSC of 88%.
73
Khan proposed generalized focal loss based on the Tversky Lou et al. achieved a DSC of 90% by introducing inverted
49
74
index for the attention UNet and achieved a DSC of 80%. residual pyramid block and context-aware fusion block
They achieved a DSC of 66% for the UNet model with focal modules to UNet architecture. By introducing CTG-Net
Tversky loss. Zhuang et al. proposed a Residual-Dilated- that integrates lesion segmentation and tumor classification
50
Attention-Gate-UNet model obtaining a DSC of 85% and tasks in breast US image analysis, Yang et al. achieved
75
reported a DSC of 82% for UNet model. Costa et al. improved performance compared to existing multi-task
51
proposed FCN-based segmentation models and reported a learning approaches. Table 2 summarizes previous works
DSC of 82%. Liang et al. developed a multi-stage elastic that utilized Dataset_A.
52
augmentation technique and achieved a DSC of 84% using a
Mask-RCNN-based segmentation network. 53
Table 2. Summary of previous works and their reported DSC
Amiri et al. developed a two-stage segmentation scores on Dataset_A
54
UNet to first detect the tumor region and then segment Article Method DSC (%)
the detected region. They reported a DSC of 86%. Lee
et al. proposed an attention module and obtained a DSC Yap et al. 47 Pre-trained model 55
55
of 76%. Shareef et al. proposed Small Tumor-Aware Abraham and Khan 49 Tversky focal loss 80
56
Network (STAN) that involved CNN layers with various Zhuang et al. 50 RDAU-Net 85
kernel sizes in order to extract multi-scale information Costa et al. 51 FCN-based model 82
from US images. They achieved a DSC of 78%. In their Liang et al. 52 Multi-stage AUG 84
next study, they improved their work by proposing an Amiri et al. 54 Two-stage UNet 86
57
enhanced STAN network and achieved a DSC of 82%. 55
Singh et al. proposed a contextual information-aware Lee et al. Attention module 76
58
network based on conditional generative adversarial Shareef et al. 56 STAN model 78
networks that integrates atrous-convolution, channel Shareef et al. 57 ESTAN model 82
60
59
attention and channel weighting, obtaining a DSC of Singh et al. 58 cGAN-based model 86
61
86%. A methodology built on the combination of deep Hussain et al. 63 DL+LS framework 98 (Benign)
62
learning (i.e., UNet network) and a traditional learning- 72 (Malignant)
based algorithm (i.e., level-set framework), proposed by Qu et al. 64 ASFRRN model 84
Hussain et al., was reported to yield the DSC of only 66
63
98% and 72% for benign and malignant tumors. Qu et al. Ning et al. Coarse-to-fine fusion 85
64
introduced an attention-supervised full-resolution residual Behboodi et al. 67 Pre-trained model 57
network inspired from full-resolution residual networks Gao et al. 68 MS fused model 85
65
and achieved a DSC of 84%. Ning et al. achieved a DSC Su et al. 69 MS UNet 82
66
of 85% from their proposed coarse-to-fine fusion network Xu et al. 70 MS self-attention model 83
alongside a weighted-balanced loss function. In one of our Huang et al. 71 Transfer learning 83
previous works, we explored the different pre-training 72
67
strategies for training a UNet when only 20 images were Yeung et al. Unified focal loss 82
used for training and obtained a maximum DSC of 57%. Xu et al. 73 Adaptive RF model 88
Lou et al. 74 IRPB+CFB modules 90
Gao et al. investigated class imbalance in segmentation
68
by proposing their multi-scale fused network with Yang et al. 75 CTG-Net 79
additive channel-spatial attention and achieved a DSC Lee et al. 44 TTFT KD-based 89
of 85%. Su et al. proposed a multi-scale UNet that Abbreviations: Aug: Augmentation; DL+LS: Deep
69
involves layers with different receptive fields and led to learning+level-set; MS: Multi-scale; RF: Receptive field;
a DSC of 82%. Xu et al. introduced a multi-scale self- RDAU-Net: Residual-Dilated-Attention-Gate-UNet; cGAN: Conditional
70
generative adversarial networks; ASFRRN: Attention-supervised
attention network by integrating local features and global full-resolution residual network; IRPB: Inverted residual pyramid block;
contextual information that led to a DSC of 83%. Huang CFB: Context-aware fusion block.
Volume 2 Issue 2 (2025) 76 doi: 10.36922/aih.3509