Dermoscopic image classification based on multi-scale and three-dimensional interaction feature optimization

Stage	Output size	MTIFNet
Average pool，7-d fc，softmax
Conv1	$\frac{H}{2} \times \frac{H}{2}$	Conv 3×3 ×3
Maxpool	$\frac{H}{4} \times \frac{H}{4}$	Conv 3×3
MTIFNet Stage1	$\frac{H}{4} \times \frac{H}{4}$	$[\begin{array}{l} C o n v 1 \times 1 \\ T I F O M \\ C o n v 3 \times 3 \\ C o n v 1 \times 1 \end{array}] \times 3$
MTIFNet Stage2	$\frac{H}{8} \times \frac{H}{8}$	$[\begin{array}{l} C o n v 1 \times 1 \\ T I F O M \\ C o n v 3 \times 3 \\ C o n v 1 \times 1 \end{array}] \times 4$
MSAM	$\frac{H}{8} \times \frac{H}{8}$	$[\begin{array}{l} S A F R M \\ D W C C M \end{array}]$
MTIFNet Stage3	$\frac{H}{16} \times \frac{H}{16}$	$[\begin{array}{l} C o n v 1 \times 1 \\ T I F O M \\ C o n v 3 \times 3 \\ C o n v 1 \times 1 \end{array}] \times 23$
MTIFNet Stage4	$\frac{H}{32} \times \frac{H}{32}$	$[\begin{array}{l} C o n v 1 \times 1 \\ T I F O M \\ C o n v 3 \times 3 \\ C o n v 1 \times 1 \end{array}] \times 3$

Table 2. Comparison of effects of different models using MSAM

View table

View in Article

Table 2. Comparison of effects of different models using MSAM

Method	MSAM	Acc	Pre	Rec	Spe	FLOPs/G	Time/s
Densenet121		85.57±0.47	76.35±0.93	72.15±0.56	96.42±0.47	239.55	0.560 7
Densenet121	√	86.71±0.41	76.93±0.58	72.10±0.57	97.02±0.43	265.34	0.575 4
Fcanet50		90.71±0.35	83.47±0.63	86.79±0.93	97.36±0.34	43.03	0.568 7
Fcanet50	√	91.11±0.61	86.86±0.62	86.90±2.11	97.73±0.78	249.93	0.821 6
ResNet101		90.30±0.65	84.08±0.90	84.56±0.40	97.44±0.49	81.84	0.722 7
ResNet101	√	90.65±0.62	85.63±0.61	84.64±0.93	97.90±0.46	94.79	0.577 1
ConvNext		92.07±0.46	88.00±0.96	88.04±0.49	97.97±0.36	160.67	0.583 4
ConvNext	√	92.95±0.50	90.39±0.79	87.98±0.88	97.99±1.01	163.90	0.858 7
ResNeSt101		92.56±0.35	89.02±0.96	86.33±1.92	98.00±0.44	107.04	0.575 5
ResNeSt101	√	93.85±0.56	91.23±2.53	90.70±2.65	98.27±0.18	119.98	0.609 0

Table 3. Ablation experiment of MSAM position

View table

View in Article

Table 3. Ablation experiment of MSAM position

Method	Acc	Pre	Rec	Spe	FLOPs/G	Time/s
Stage1	92.66±0.32	88.92±0.51	85.69±1.58	97.96±0.24	120.04	0.578 7
Stage2	93.85±0.56	91.23±2.53	90.70±2.65	98.27±0.18	119.98	0.609 0
Stage3	92.45±0.41	89.29±3.30	85.77±1.37	97.87±0..15	119.96	0.671 9
Stage4	92.70±0.31	88.68±1.49	86.05±1.19	98.07±0.39	119.94	0.810 2

Table 4. Validation of TIFOM

View table

View in Article

Table 4. Validation of TIFOM

Method			Acc	Pre	Rec	Spe	FLOPs/G	Time/s
Base	Triplet Attention	SimAM Attention	Acc	Pre	Rec	Spe	FLOPs/G	Time/s
√			92.56±0.35	89.02±0.96	86.33±1.92	98.00±0.44	107.04	0.575 5
√	√		93.00±0.51	89.18±1.70	87.03±0.96	98.10±0.11	107.30	0.807 9
√		√	93.01±0.55	89.35±1.18	87.15±1.75	97.97±0.39	107.04	0.876 7
√	√	√	93.51±0.52	89.88±1.33	89.57±2.67	98.20±0.21	107.30	0.811 9

Table 5. Ablation experiments of the whole network

View table

View in Article

Table 5. Ablation experiments of the whole network

Method				Acc	Pre	Rec	Spe	FLOPs/G	Time/s
Base	SAFRM	DWCCM	TIFOM	Acc	Pre	Rec	Spe	FLOPs/G	Time/s
√				92.56±0.35	89.02±0.96	86.33±1.92	98.00±0.44	107.04	0.575 5
√	√			92.94±0.35	89.70±1.09	89.70±3.27	97.95±0..16	109.21	0.526 5
√	√	√		93.85±0.56	91.23±2.53	90.70±2.65	98.27±0.18	119.98	0.609 0
√			√	93.51±0.52	89.88±1.33	89.57±2.67	98.20±0.21	107.30	0.811 9
√	√	√	√	94.32±0.45	91.61±1.29	93.00±0.60	98.39±0.17	120.24	0.937 3

Table 6. Classification results of different models on the ISIC 2018 dataset
View table
View in Article
Table 6. Classification results of different models on the ISIC 2018 dataset
Method Acc Pre Rec Spe
Le^［19］ 93.00 88.00 86.00 -
Liu^［20］ 93.14 47.70 81.84 92.30
Datta^［21］ 93.40 93.70 - -
Cai^［22］ 93.81 88.64 90.14 98.36
Yang^［23］ 94.10 82.43 84.43 97.86
Ours 94.32 91.61 93.00 98.39

Table 7. Classification results of different models on the ISIC 2017 dataset
View table
View in Article
Table 7. Classification results of different models on the ISIC 2017 dataset
Method Acc Pre Rec Spe
Zhang^［24］ 86.80 - 87.80 86.70
Al-Masni^［25］ 81.57 - 75.33 80.62
Datta^［21］ 90.40 - 91.60 83.30
Kim^［29］ 89.50 91.30 49.60 99.20
Khouloud^［26］ 96.97 95.71 95.15 97.87
Ours 98.57 98.20 98.47 99.13

Tools

Get Citation

Copy Citation Text

Di WANG, Xiaoqi LÜ, Jing LI. Dermoscopic image classification based on multi-scale and three-dimensional interaction feature optimization[J]. Optics and Precision Engineering, 2024, 32(24): 3644

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites